The Cylc Suite Engine
User Guide
6.10.0
GNU GPL v3.0 Software License
Copyright (C) 2008-2016 NIWA
May 5, 2016
Contents
Environmental forecasting suites generate forecast products from a potentially large group of interdependent scientific models and associated data processing tasks. They are constrained by availability of external driving data: typically one or more tasks will wait on real time observations and/or model data from an external system, and these will drive other downstream tasks, and so on. The dependency diagram for a single forecast cycle point in such a system is a Directed Acyclic Graph as shown in Figure 1 (in our terminology, a forecast cycle point is comprised of all tasks with a common cycle point, which is the nominal analysis time or start time of the forecast models in the group). In real time operation processing will consist of a series of distinct forecast cycle points that are each initiated, after a gap, by arrival of the new cycle point’s external driving data.
From a job scheduling perspective task execution order in such a system must be carefully controlled in order to avoid dependency violations. Ideally, each task should be queued for execution at the instant its last prerequisite is satisfied; this is the best that can be done even if queued tasks are not able to execute immediately because of resource contention.
Cylc was developed for the EcoConnect Forecasting System at NIWA (National Institute of Water and Atmospheric Research, New Zealand). EcoConnect takes real time atmospheric and stream flow observations, and operational global weather forecasts from the Met Office (UK), and uses these to drive global sea state and regional data assimilating weather models, which in turn drive regional sea state, storm surge, and catchment river models, plus tide prediction, and a large number of associated data collection, quality control, preprocessing, post-processing, product generation, and archiving tasks.1 The global sea state forecast runs once daily. The regional weather forecast runs four times daily but it supplies surface winds and pressure to several downstream models that run only twice daily, and precipitation accumulations to catchment river models that run on an hourly cycle assimilating real time stream flow observations and using the most recently available regional weather forecast. EcoConnect runs on heterogeneous distributed hardware, including a massively parallel supercomputer and several Linux servers.
Most dependence between tasks applies within a single forecast cycle point. Figure 1 shows the dependency diagram for a single forecast cycle point of a simple example suite of three forecast models (a, b, and c) and three post processing or product generation tasks (d, e and f). A scheduler capable of handling this must manage, within a single forecast cycle point, multiple parallel streams of execution that branch when one task generates output for several downstream tasks, and merge when one task takes input from several upstream tasks.
Figure 2 shows the optimal job schedule for two consecutive cycle points of the example suite in real time operation, given execution times represented by the horizontal extent of the task bars. There is a time gap between cycle points as the suite waits on new external driving data. Each task in the example suite happens to trigger off upstream tasks finishing, rather than off any intermediate output or event; this is merely a simplification that makes for clearer diagrams.
Now the question arises, what happens if the external driving data for upcoming cycle points is available in advance, as it would be after a significant delay in operations, or when running a historical case study? While the forecast model a appears to depend only on the external data x at this stage of the discussion, in fact it would typically also depend on its own previous instance for the model background state used in initializing the new forecast. Thus, as alluded to in Figure 3, task a could in principle start as soon as its predecessor has finished. Figure 4 shows, however, that starting a whole new cycle point at this point is dangerous - it results in dependency violations in half of the tasks in the example suite. In fact the situation could be even worse than this - imagine that task b in the first cycle point is delayed for some reason after the second cycle point has been launched. Clearly we must consider handling inter-cycle dependence explicitly or else agree not to start the next cycle point early, as is illustrated in Figure 5.
Forecast models typically depend on their own most recent previous forecast for background state or restart files of some kind (this is called warm cycling) but there can also be inter-cycle dependence between different tasks. In an atmospheric forecast analysis suite, for instance, the weather model may generate background states for observation processing and data-assimilation tasks in the next cycle point as well as for the next forecast model run. In real time operation inter-cycle dependence can be ignored because it is automatically satisfied when one cycle point finishes before the next begins. If it is not ignored it drastically complicates the dependency graph by blurring the clean boundary between cycle points. Figure 6 illustrates the problem for our simple example suite assuming minimal inter-cycle dependence: the warm cycled models (a, b, and c) each depend on their own previous instances.
For this reason, and because we tend to see forecasting suites in terms of their real time characteristics, other metaschedulers have ignored inter-cycle dependence and are thus restricted to running entire cycle points in sequence at all times. This does not affect normal real time operation but it can be a serious impediment when advance availability of external driving data makes it possible, in principle, to run some tasks from upcoming cycle points before the current cycle point is finished - as was suggested at the end of the previous section. This can occur, for instance, after operational delays (late arrival of external data, system maintenance, etc.) and to an even greater extent in historical case studies and parallel test suites started behind a real time operation. It can be a serious problem for suites that have little downtime between forecast cycle points and therefore take many cycle points to catch up after a delay. Without taking account of inter-cycle dependence, the best that can be done, in general, is to reduce the gap between cycle points to zero as shown in Figure 5. A limited crude overlap of the single cycle point job schedule may be possible for specific task sets but the allowable overlap may change if new tasks are added, and it is still dangerous: it amounts to running different parts of a dependent system as if they were not dependent and as such it cannot be guaranteed that some unforeseen delay in one cycle point, after the next cycle point has begun, (e.g. due to resource contention or task failures) won’t result in dependency violations.
Figure 7 shows, in contrast to Figure 4, the optimal two cycle point job schedule obtained by respecting all inter-cycle dependence. This assumes no delays due to resource contention or otherwise - i.e. every task runs as soon as it is ready to run. The scheduler running this suite must be able to adapt dynamically to external conditions that impact on multi-cycle-point scheduling in the presence of inter-cycle dependence or else, again, risk bringing the system down with dependency violations.
To further illustrate the potential benefits of proper inter-cycle dependency handling, Figure 8 shows an operational delay of almost one whole cycle point in a suite with little downtime between cycle points. Above the time axis is the optimal schedule that is possible in principle when inter-cycle dependence is taken into account, and below it is the only safe schedule possible in general when it is ignored. In the former case, even the cycle point immediately after the delay is hardly affected, and subsequent cycle points are all on time, whilst in the latter case it takes five full cycle points to catch up to normal real time operation.
Similarly, Figure 9 shows example suite job schedules for an historical case study, or when catching up after a very long delay; i.e. when the external driving data are available many cycle points in advance. Task a, which as the most upstream forecast model is likely to be a resource intensive atmosphere or ocean model, has no upstream dependence on co-temporal tasks and can therefore run continuously, regardless of how much downstream processing is yet to be completed in its own, or any previous, forecast cycle point (actually, task a does depend on co-temporal task x which waits on the external driving data, but that returns immediately when the data is available in advance, so the result stands). The other forecast models can also cycle continuously or with a short gap between, and some post processing tasks, which have no previous-instance dependence, can run continuously or even overlap (e.g. e in this case). Thus, even for this very simple example suite, tasks from three or four different cycle points can in principle run simultaneously at any given time.
In fact, if our tasks are able to trigger off internal outputs of upstream tasks (message triggers) rather than waiting on full completion, then successive instances of the forecast models could overlap as well (because model restart outputs are generally completed early in the forecast) for an even more efficient job schedule.
Cylc manages a pool of proxy objects that represent the real tasks in a suite. Task proxies know how to run the real tasks that they represent, and they receive progress messages from the tasks as they run (usually reports of completed outputs). There is no global cycling mechanism to advance the suite; instead individual task proxies have their own private cycle point and spawn their own successors when the time is right. Task proxies are self-contained - they know their own prerequisites and outputs but are not aware of the wider suite. Inter-cycle dependence is not treated as special, and the task pool can be populated with tasks with many different cycle points. The task pool is illustrated in Figure 10. Whenever any task changes state due to completion of an output, every task checks to see if its own prerequisites have been satisfied. In effect, cylc gets a pool of tasks to self-organize by negotiating their own dependencies so that optimal scheduling, as described in the previous section, emerges naturally at run time.
The following packages are technically optional as you can construct and run cylc suites without dependency graphing, the gcylc GUI, or template processing but this is not recommended, and without Jinja2 you will not be able to run many of the example suites:
If you use a binary package manager to install graphviz you may also need a couple of devel packages for the pygraphviz build:
This user guide can be generated from the LATEXsource by running make in the top level cylc directory after download. The following TEXpackages are required (but note that the exact packages required may be somewhat OS or distribution-dependent):
And for HTML versions of the User Guide:
Note: Older versions of cylc require Pyro-3 (Python Remote Objects). Current version of cylc includes a
modified version of Pyro 3.16 in its distribution. Pyro is used for network communication between
server processes (cylc suites) and client programs (running tasks, the gcylc GUI, user commands).
https://pypi.python.org/pypi/Pyro/
Finally, cylc makes heavy use of Python ordered dictionary data structures. Significant speedup in parsing large suites can be had by installing the fast C-coded ordereddict module by Anthon van der Neut:
This module is currently included with cylc under $CYLC_DIR/ext, and is built by the top level cylc Makefile. If you install the resulting library appropriately cylc will automatically use it in place of a slower Python implementation of the ordered dictionary structure.
Cylc should run “out of the box” on recent Linux distributions.
It has been reported that cylc runs fine on OSX 10.6 SnowLeopard, but on OSX 10.7 Lion there is an issue with constructing proper FQDNs (Fully Qualified Domain Names) that requires a change to the DNS service. Here’s how to solve the problem:
Cylc has incorporated a custom-modified version the xdot graph viewer (https://github.com/jrfonseca/xdot.py, LGPL license).
First install graphviz, Pygraphviz, Jinja2, TEX, and ImageMagick using the package manager on your system if possible; otherwise download the packages manually and follow their native installation documentation. On a modern Linux system, this is very easy. For example, to install cylc-5.1.0 on the Fedora 18 Linux distribution:
If you do not have root access on your intended cylc host machine and cannot get a sysadmin to do this at system level, see 4.4 for tips on installing everything to a local user account.
Now check that everything other than the LATEXpackages is installed properly:
If this command reports any errors then the packages concerned are not installed, not in the system Python search path, or (for a local install) not present in your $PYTHONPATH variable.
Cylc installs into a normal user account, as an unpacked release tarball or a git repository clone. See the INSTALL file in the source tree for instructions (also listed in G).
Site and user global config files define some important parameters that affect all suites, some of which may need to be customized for your site. See 6 for how to generate an initial site file and where to install it. All legal site and user global config items are defined in B.
If your users submit task jobs to hosts other than the hosts they use to run their suites, you should ensure that the job hosts have the correct environment for running cylc. A cylc suite generates task job scripts that normally invoke bash. The job will attempt to source the first of these files it finds to set up its environment:
The ${CYLC_DIR}/conf/job-init-env-default.sh file is provided in the cylc distribution, and will attempt to source /etc/profile and ${HOME}/.profile. If this behaviour is not desirable, you should override it by adding a ${CYLC_DIR}/conf/job-init-env.sh file and populate it with the appropriate contents.
The cylc test battery is primarily intended for developers to check that changes to the source code don’t break existing functionality. Note that some test failures can be expected to result from suites timing out, even if nothing is wrong, if you run too many tests in parallel. See cylc test-battery --help.
It is possible to install cylc and all of its software prerequisites under your own user account. Cylc itself is already designed to be installed into a normal user account, just follow the instructions above in 4.2. For the other packages, depending on the installation method used for each, it is just a matter of learning how to change the default install path prefix from, for example, /usr/local to $HOME/installed/usr/local and then ensuring that the resulting local package paths are set properly in your PYTHONPATH environment variable.
The graphviz build reportedly may fail on systems that do not have QT installed, hence the ./configure --with-qt=no option above. The graphviz lib and include locations are required when installing Pygraphviz.
Finally, check that everything (other than LATEXfor document processing) is installed:
If this command reports any errors then the packages concerned are not installed, not in the system Python search path, or (for a local install) not present in your $PYTHONPATH variable.
Upgrading is just a matter of unpacking the new cylc release. Successive cylc releases can be installed in parallel as suggested in the INSTALL file (G).
The problem with the word “cycle” is that it is often used to refer to a series of times (e.g. “an hourly cycle”) as well as a particular time (e.g. “current cycle” or “current cycle point”). Here, we use “cycle point” to refer to a particular (usually date-time) item. If a task has an hourly date-time cycle that starts at 2014-01-01T00 and ends at 2014-01-02T00, the first “cycle point” will be 2014-01-01T01, the second “cycle point” will be 2014-01-01T02, etc.
You may be accustomed to the idea that a forecasting suite has a “current cycle” (our nomenclature: “current cycle point”), which is typically the analysis time or nominal start time of the main forecast model(s) in the suite, and that the whole suite advances to the next forecast cycle point when all tasks in the current cycle point have finished (or even when a particular wall clock time is reached, in real time operation). As explained in the Introduction, this is not how cylc works.
Cylc suites advance by means of individual tasks with private cycle points (times) independently spawning successors at the next valid cycle point for the task, not by incrementing a suite-wide forecast cycle point. Each task will be submitted when its own prerequisites are satisfied, regardless of other tasks with other cycle points running, or not, at the time. It may still be convenient at times, however, to refer to the “current cycle point”, the “previous cycle point”, or the “next cycle point” and so forth, with reference to a particular task, or in the sense of all tasks that “belong to” a particular forecast cycle point. But keep in mind that the members of these groups may not be present simultaneously in the running suite - i.e. different tasks may pass through the “current cycle point” (etc.) at different times as the suite evolves, particularly in delayed (catch up) operation.
Cylc site and user global configuration files contain settings that affect all suites. Some of these, such as the range of network ports used by cylc, should be set at site level,
Others, such as the preferred text editor for suite definitions, can be overridden by users,
The cylc get-site-config command retrieves current global settings consisting of cylc defaults overridden by site settings, if any, overridden by user settings, if any. To generate an initial site or user global config file:
Settings that do not need to be changed should be deleted or commented out of user global config files so that they don’t override future changes to the site file.
Legal items, values, and system defaults are documented in (B).
This section provides a hands-on tutorial introduction to basic cylc suite preparation and control. A number of features are not yet touched on by the tutorial examples, however, so please also read the rest of the User Guide.
Some global parameters affecting cylc’s behaviour are defined in a site global config file, and can be customized per user in user global config files. For example, to choose the text editor invoked by cylc on suite definitions:
If you submit task jobs to hosts other than the hosts you use to run your suites, you may need to customise the environment for running cylc. A cylc suite generates task job scripts that normally invoke bash. The job will attempt to source the first of these files it finds to set up its environment:
The ${CYLC_DIR}/conf/job-init-env-default.sh file is provided in the cylc distribution, and will attempt to source /etc/profile and ${HOME}/.profile. If this behaviour is not desirable, your site administrator who installed cylc should have overridden it by adding a ${CYLC_DIR}/conf/job-init-env.sh file and populate it with the appropriate contents. If customisation is still required, you can add your own ${HOME}/.cylc/job-init-env.sh file and populate it with the appropriate contents.
Cylc has command line (CLI) and graphical (GUI) user interfaces. To access them do not simply put the bin directory of the current release the shell search path for users. Instead, as described in the cylc INSTALL file (G), configure the provided multi-version wrapper script to point to install location for any or all cylc releases. This allows existing suites to continue running at old cylc versions if necessary.
The command line interface is unified under a single top level cylc command that provides access to many sub-commands and their help documentation.
The cylc GUI covers the same functionality as the CLI, but it has more sophisticated suite monitoring capability. It can start and stop suites, or connect to suites that are already running; in either case, shutting down the GUI does not affect the suite itself.
Clicking on a suite in the scan GUI, shown in Figure 13, opens a gcylc instance for it.
Cylc suites are defined by extended-INI format suite.rc files (the main file format extension is section nesting). These reside in suite definition directories that may also contain a bin directory and any other suite-related files.
Suite registration associates a name with a suite definition directory, in a simple database. Cylc commands that parse suite definition files can take the file path or the suite name as input; commands that interact with running suites have to target the suite by name.
A suite-specific passphrase file is automatically generated in the suite definition directory at registration time. It is loaded by the suite daemon at start-up and used to authenticate all client connections. Clients on the suite host account automatically load the passphrase from the suite definition directory. The passphrase is also installed automatically in the suite run directory on hosts for running task jobs. On other accounts that do not share the same HOME directory with the suite host account and do not have non-interactive SSH access to the suite host account, it is possible to install the passphrase manually. (A client will attempt to install the passphrase automatically via non-interactive SSH.)
Possession of a suite passphrase gives full control and full read-access to the suite. Without it, client access is determined by the suite’s public access privilege level.
For more on connection authentication, suite passphrases, and public access, see 12.6.
Run the following command to import cylc’s example suites to a chosen directory location and register them for use under the examples name group:
(first check that $TMPDIR is defined in your environment, or else use a different location). List the newly registered tutorial suites using the cylc print command:
See cylc db print --help for other display options. The tree-form display shows how hierarchical suite names can be used to organize related suites nicely (suite names do not have to be related to their source directory paths, although they are in this case):
Suite definitions can be validated against the suite.rc file format specification to detect many types of error without running the suite.
Here’s the traditional Hello World program rendered as a cylc suite:
Cylc suites feature a clean separation of scheduling configuration, which determines when tasks are ready to run; and runtime configuration, which determines what to run (and where and how to run it) when a task is ready. In this example the [scheduling] section defines a single task called hello that triggers immediately when the suite starts up. When the task finishes the suite shuts down. That this is a dependency graph will be more obvious when more tasks are added. Under the [runtime] section the script item defines a simple inlined implementation for hello: it sleeps for ten seconds, then prints Hello World!, and exits. This ends up in a job script generated by cylc to encapsulate the task (below) and, thanks to some defaults designed to allow quick prototyping of new suites, it is submitted to run as a background job on the suite host. In fact cylc even provides a default task implementation that makes the entire [runtime] section technically optional:
(the resulting dummy task just prints out some identifying information and exits).
The text editor invoked by cylc on suite definitions is determined by cylc site and user global config files, as shown above in 7.2. Check that you have renamed the tutorial examples suites as described just above and open the Hello World suite definition in your text editor:
Alternatively, start gcylc on the suite,
and choose Suite →Edit from the menu.
The editor will be invoked from within the suite definition directory for easy access to other suite files (in this case there are none). There are syntax highlighting control files for several text editors under /path/to/cylc/conf/; see in-file comments for installation instructions.
If you’re quick enough (this example only takes 10-15 seconds to run) the cylc scan command will detect the running suite:
Using the --no-detach option prevents the suite daemonizing, but only warnings and errors, if any, are printed to the terminal. The --debug option, however, causes (among other things) task job submission commands to to be printed stdout:
When a task is ready to run a job script is generated to run it. The command printed just above is how cylc executes background jobs on the suite host (this is the default), retrieves the job process ID, and directs the output to log files in standard locations. For the log file locations that can be seen above, in job/1/hello/01/ the first 1/ is the cycle point of the task hello (for a non-cycling task this is just the integer 1); and the final 01/ directory is its submission number (tasks can be made to retry on failure, manually or automatically, in which case the submission number increments).
The suite shuts down automatically once all tasks have succeeded.
The cylc GUI can start and stop suites, or (re)connect to suites that are already running:
Use the tool bar Play button, or the Control →Run menu item, to run the suite again. You may want to alter the suite definition slightly to make the task take longer to run. Try right-clicking on the hello task to view its output logs. The relative merits of the three suite views - dot, text, and graph - will be more apparent later when we have more tasks. Closing the GUI does not affect the suite itself.
Suites that are currently running can be detected with command line or GUI tools:
The scan GUI is shown in Figure 13; clicking on a suite in it opens gcylc.
At run time, task instances are identified by name, which is determined entirely by the suite definition, and a cycle point which is usually a date-time or an integer:
Non-cycling tasks usually just have the cycle point 1, but this still has to be used to target the task instance with cylc commands.
Task job scripts are generated by cylc to wrap the task implementation specified in the suite definition (environment, script, etc.) in error trapping code, messaging calls to report task progress back to the suite daemon, and so forth. Job scripts are written to the suite job log directory where they can be viewed alongside the job output logs. They can be accessed at run time by right-clicking on the task in the cylc GUI, or printed to the terminal:
This command can also print the suite log (and stdout and stderr for suites in daemon mode) and task stdout and stderr logs (see cylc log --help). A new job script can also be generated on the fly for inspection,
Take a look at the job script generated for hello.1 during the suite run above. The custom scripting should be clearly visible toward the bottom of the file.
The hello task in the first tutorial suite defaults to running as a background job on the suite host. To submit it to the Unix at scheduler instead, configure its job submission settings as in tut.oneoff.jobsub:
If you run the suite (first check that the at daemon atd is running on the suite host) you’ll see a slightly different job submission command printed to stdout:
Cylc supports a number of different job submission methods. Tasks submitted to external batch queuing systems like at, PBS, SLURM, Moab, or LoadLeveler, are displayed as submitted in the cylc GUI until they actually start executing.
If the --no-detach option is not used, suite stdout and stderr will be directed to the suite run directory along with the time-stamped suite log file, and task job scripts and job logs (task stdout and stderr). The default suite run directory location is $HOME/cylc-run:
The suite run database, suite environment file, suite state files, and task status files are used internally by cylc. Tasks execute in private work/ directories that are deleted automatically if empty when the task finishes. The suite share/ directory is made available to all tasks (by $CYLC_SUITE_SHARE_DIR) as a common share space. The task submission number increments from 1 if a task retries on failure; this is used a sub-directory of the log tree to avoid overwriting log files from earlier job submissions.
The top level run directory location can be changed in site and user config files if necessary, and the suite share and work locations can be configured separately because of the potentially larger disk space requirement.
Task job logs can be viewed by right-clicking on tasks in the gcylc GUI (so long as the task proxy is live in the suite), manually accessed from the log directory (of course), or printed to the terminal with the cylc log command:
The hello task in the first two tutorial suites defaults to running on the suite host. To make it run on a remote host instead change its runtime configuration as in tut.oneoff.remote:
For remote task hosting to work several requirements must be satisfied:
If your username is different on the task host the [[[remote]]] section also supports an owner=username item, or your $HOME/.ssh/config file can be configured for username translation.
If you configure a task host according to the requirements above and run the suite again in debug mode, you’ll see that the job submission command printed to suite stdout is now considerably more complicated. That’s because it has to create remote log directories, source login scripts on the remote host to ensure cylc is visible there, send the task job script over, and submit it to run there by the configured job submission method:
(Don’t be intimidated by this - it is really quite straightforward and would appear much simpler if the job log path was shorter!). Remote task job logs are saved to the suite run directory on the task host, not on the suite host, although they can be retrieved by right-clicking on the task in the GUI. If you want the job logs pulled back to the suite host automatically, you can set ”retrieve job logs=True” under the ”[[[remote]]]” section:
The suite will attempt to rsync the job logs once from the remote host each time a task job completes. E.g. if the job file is ~/cylc-run/tut.oneoff.remote/log/job/1/hello/01/job, anything under ~/cylc-run/tut.oneoff.remote/log/job/1/hello/01/ will be retrieved.
Some batch systems have considerable delays between the time when the job completes and when it writes the job logs in its normal location. If this is the case, you can configure an initial delay and retry delays for job log retrieval by setting some delays. E.g.:
Finally, if the disk space of the suite host is limited, you may want to set [[[remote]]]retrieve job logs max size=SIZE. The value of SIZE can be anything that is accepted by the --max-size=SIZE option of the rsync command. E.g.:
It is worth noting that cylc uses the existence of a job’s job.out or job.err in the local file system to indicate a successful job log retrieval. If retrieve job logs max sizeSIZE= is set and both job.out and job.err are bigger than SIZE then cylc will consider the retrieval as failed. If retry delays are specified, this will trigger some useless (but harmless) retries. If this occurs regularly, you should try the following:
To make a second task called goodbye trigger after hello finishes successfully, return to the original example, tut.oneoff.basic, and change the suite graph as in tut.oneoff.goodbye:
or to trigger it at the same time as hello,
and configure the new task’s behaviour under [runtime]:
Run tut.oneoff.goodbye and check the output from the new task:
Task names in the graph string can be qualified with a state indicator to trigger off task states other than success:
A common use of this is to automate recovery from known modes of failure:
i.e. if task goodbye fails, trigger another task that (presumably) really says goodbye.
Failure triggering generally requires use of suicide triggers as well, to remove the recovery task if it isn’t required (otherwise it would hang about indefinitely in the waiting state):
This means if goodbye fails, trigger really_goodbye; and otherwise, if goodbye succeeds, remove really_goodbye from the suite.
Try running tut.oneoff.suicide, which also configures the hello task’s runtime to make it fail, to see how this works.
The [runtime] section is actually a multiple inheritance hierarchy. Each subsection is a namespace that represents a task, or if it is inherited by other namespaces, a family. This allows common configuration to be factored out of related tasks very efficiently.
The [root] namespace provides defaults for all tasks in the suite. Here both tasks inherit script from root, which they customize with different values of the environment variable $GREETING. Note that inheritance from root is implicit; from other parents an explicit inherit = PARENT is required, as shown below.
Task families defined by runtime inheritance can also be used as shorthand in graph trigger expressions. To see this, consider two “greeter” tasks that trigger off another task foo,
If we put the common greeting functionality of greeter_1 and greeter_2 into a special GREETERS family, the graph can be expressed more efficiently like this:
i.e. if foo succeeds, trigger all members of GREETERS at once. Here’s the full suite with runtime hierarchy shown:
(Note that we recommend given ALL-CAPS names to task families to help distinguish them from task names. However, this is just a convention).
Verbose validation shows the family member substitution done when the suite definition is parsed:
Experiment with the tut.oneoff.ftrigger1 suite to see how this works.
Tasks (or families) can also trigger off other families, but in this case we need to specify what the trigger means in terms of the upstream family members. Here’s how to trigger another task bar if all members of GREETERS succeed:
Verbose validation in this case reports:
Cylc ignores family member qualifiers like succeed-all on the right side of a trigger arrow, where they don’t make sense, to allow the two graph lines above to be combined in simple cases:
Any task triggering status qualified by -all or -any, for the members, can be used with a family trigger. For example, here’s how to trigger bar if all members of GREETERS finish (succeed or fail) and any of them them succeed:
(use of GREETERS:succeed-any by itself here would trigger bar as soon as any one member of GREETERS completed successfully). Verbose validation now begins to show how family triggers can simplify complex graphs, even for this tiny two-member family:
Experiment with tut.oneoff.ftrigger2 to see how this works.
You can style dependency graphs with an optional [visualization] section, as shown in tut.oneoff.ftrigger2:
To display the graph in an interactive viewer,
It should look like Figure 15 (with the GREETERS family node expanded on the right).
Graph styling can be applied to entire families at once, and custom “node groups” can also be defined for non-family groups.
The tasks in our examples so far have all had inlined implementation, in the suite definition, but real tasks often need to call external commands, scripts, or executables. To try this, let’s return to the basic Hello World suite and cut the implementation of the task hello out to a file hello.sh in the suite bin directory:
Make the task script executable, and change the hello task runtime section to invoke it:
If you run the suite now the new greeting from the external task script should appear in the hello task stdout log. This works because cylc automatically adds the suite bin directory to $PATH in the environment passed to tasks via their job scripts. To execute scripts (etc.) located elsewhere you can refer to the file by its full file path, or set $PATH appropriately yourself (this could be done via $HOME/.profile, which is sourced at the top of the task job script, or in the suite definition itself).
Note the use of set -e above to make the script abort on error. This allows the error trapping code in the task job script to automatically detect unforeseen errors.
So far we’ve considered non-cycling tasks, which finish without spawning a successor.
Cycling is based around iterating through date-time or integer sequences. A cycling task may run at each cycle point in a given sequence (cycle). For example, a sequence might be a set of date-times every 6 hours starting from a particular date-time. A cycling task may run for each date-time item (cycle point) in that sequence.
There may be multiple instances of this type of task running in parallel, if the opportunity arises and their dependencies allow it. Alternatively, a sequence can be defined with only one valid cycle point - in that case, a task belonging to that sequence may only run once.
Open the tut.cycling.one suite:
The difference between cycling and non-cycling suites is all in the [scheduling] section, so we will leave the [runtime] section alone for now (this will result in cycling dummy tasks). Note that the graph is now defined under a new section heading that makes each task under it have a succession of cycle points ending in 00 or 12 hours, between specified initial and final cycle points (or indefinitely if no final cycle point is given), as shown in Figure 16.
If you run this suite instances of foo will spawn in parallel out to the runahead limit, and each bar will trigger off the corresponding instance of foo at the same cycle point. The runahead limit, which defaults to a few cycles but is configurable, prevents uncontrolled spawning of cycling tasks in suites that are not constrained by clock triggers in real time operation.
Experiment with tut.cycling.one to see how cycling tasks work.
ISO 8601 Date-Time Syntax The suite above is a very simple example of a cycling date-time workflow. More generally, cylc comprehensively supports the ISO 8601 standard for date-time instants, intervals, and sequences. Cycling graph sections can be specified using full ISO 8601 recurrence expressions, but these may be simplified by assuming context information from the suite - namely initial and final cycle points. One form of the recurrence syntax looks like Rn/start-date-time/period (Rn means repeat n times). In the example above, if the initial cycle point is always at 00 or 12 hours then [[[T00,T12]]] could be written as [[[PT12H]]], which is short for [[[R/initial-cycle-point/PT12H/]]] - i.e. repeat indefinitely every 12 hours starting at the initial cycle point. It is possible to add constraints to the suite to only allow initial cycle points at 00 or 12 hours e.g.
The tut.cycling.two suite adds inter-cycle dependence to the previous example:
For any given cycle point in the sequence defined by the cycling graph section heading, bar triggers off foo as before, but now foo triggers off its own previous instance foo[-PT12H]. Date-time offsets in inter-cycle triggers are expressed as ISO 8601 intervals (12 hours in this case). Figure 17 shows how this connects the cycling graph sections together.
Experiment with this suite to see how inter-cycle triggers work. Note that the first instance of foo, at suite start-up, will trigger immediately in spite of its inter-cycle trigger, because cylc ignores dependence on points earlier than the initial cycle point. However, the presence of an inter-cycle trigger usually implies something special has to happen at start-up. If a model depends on its own previous instance for restart files, for example, then some special process has to generate the initial set of restart files when there is no previous cycle point to do it. The following section shows one way to handle this in cylc suites.
Sometimes we want to be able to run a task at the initial cycle point, but refrain from running it in subsequent cycles. We can do this by writing an extra set of dependencies that are only valid at a single date-time cycle point. If we choose this to be the initial cycle point, these will only apply at the very start of the suite.
The cylc syntax for writing this single date-time cycle point occurrence is R1, which stands for R1/no-specified-date-time/no-specified-period. This is an adaptation of part of the ISO 8601 date-time standard’s recurrence syntax (Rn/date-time/period) with some special context information supplied by cylc for the no-specified-⋆ data.
The 1 in the R1 means repeat once. As we’ve specified no date-time, Cylc will use the initial cycle point date-time by default, which is what we want. We’ve also missed out specifying the period - this is set by cylc to a zero amount of time in this case (as it never repeats, this is not significant).
For example, in tut.cycling.three:
This is shown in Figure 18.
Note that the time zone has been set to +1300 in this case, instead of UTC (Z) as before. If no time zone or UTC mode was set, the local time zone of your machine will be used in the cycle points.
At the initial cycle point, foo will depend on foo[-PT12H] and also on prep:
Thereafter, it will just look like e.g.:
However, in our initial cycle point example, the dependence on foo.20130807T1200+13 will be ignored, because that task’s cycle point is earlier than the suite’s initial cycle point and so it cannot run. This means that the initial cycle point dependencies for foo actually look like:
Cylc can do also do integer cycling for repeating workflows that are not date-time based.
Open the tut.cycling.integer suite, which is plotted in Figure 19.
The integer cycling notation is intended to look similar to the ISO 8601 date-time notation, but it is simpler for obvious reasons. The example suite illustrates two recurrence forms, Rn/start-point/period and Rn/period/stop-point, simplified somewhat using suite context information (namely the initial and final cycle points). The first form is used to run one special task called start at start-up, and for the main cycling body of the suite; and the second form to run another special task called stop in the final two cycles. The P character denotes period (interval) just like in the date-time notation. R/1/P2 would generate the sequence of points 1,3,5,....
Cylc has built in support for the Jinja2 template processor, which allows us to embed code in suite definitions to generate the final result seen by cylc.
The tut.oneoff.jinja2 suite illustrates two common uses of Jinja2: changing suite content or structure based on the value of a logical switch; and iteratively generating dependencies and runtime configuration for groups of related tasks:
To view the result of Jinja2 processing with the Jinja2 flag MULTI set to False:
And with MULTI set to True:
Tasks can be configured to retry a number of times if they fail. An environment variable $CYLC_TASK_TRY_NUMBER increments from 1 on each successive try, and is passed to the task to allow different behaviour on the retry:
When a task with configured retries fails, its cylc task proxy goes into the retrying state until the next retry delay is up, then it resubmits. It only enters the failed state on a final definitive failure.
Experiment with tut.oneoff.retry to see how this works.
If you have read access to another user’s account (even on another host) it is possible to use cylc monitor to look at their suite’s progress without full shell access to their account. To do this, you will need to copy their suite passphrase to
(use of the host and owner names is optional here - see 12.6.1) and also retrieve the port number of the running suite from,
Once you have this information, you can run
to view the progress of their suite.
Other suite-connecting commands work in the same way too; see 12.10.
The cylc suite search tool reports pattern matches in the suite definition by line number, suite section, and file, even if the suite uses nested include-files, and by file and line number for matches in suite bin scripts:
Almost every feature of cylc can be tested quickly and easily with a simple dummy suite. You can write your own, or start from one of the example suites in /path/to/cylc/examples (see use of cylc import-examples above) - they all run “out the box” and can be copied and modified at will.
Cylc commands target suites via names registered in a suite name database located at $HOME/.cylc/REGDB/. Suite names are hierarchical like directory paths, allowing nested tree-like grouping, but use the ‘.’ character as a delimiter. This :
Suite titles held in the name database are parsed from the suite definition at the time of initial suite registration. If you change the title later use cylc db refresh to update the database.
Name groups are entirely virtual, they do not need to be explicitly created before use, and they automatically disappear if all tasks are removed from them. From the listing above, for example, to move the suite nwp.oper.region2 into the nwp.test group:
And to move nwp.test.region2 into a new group nwp.para:
Currently you cannot explicitly indicate a group name on the command line by appending a dot character. Rather, in database operations such as copy, reregister, or unregister, the identity of the source item (group or suite) is inferred from the content of the database; and if the source item is a group, so must the target be a group (or it will be, in the case of an item that will be created by the operation). This means that you cannot copy a single suite into a group that does not exist yet unless you specify the entire target suite name.
cylc db register --help shows a number of other examples.
On the command line, the ‘database’ (or ‘db’) command category contains commands to implement the aforementioned operations.
Groups of suites (at any level in the name hierarchy) can be deleted, copied, imported, and exported; as well as individual suites. To do this, just use suite group names as source and/or target for operations, as appropriate. For instance, if a group foo.bar contains the suites foo.bar.baz and foo.bar.qux, you can copy a single suite like this:
(resulting in a new suite boo); or the group like this:
(resulting in new suites boo.baz and boo.qux); or the group like this:
(resulting in new suites boo.bar.baz and boo.bar.qux). When suites are copied, the suite definition directories are copied into a directory tree, under the target directory, that reflects the suite name hierarchy. cylc copy --help has some explicit examples.
The same functionality is also available by right-clicking on suites or groups in the gcylc “Open Registered Suite” dialog.
Cylc suites are defined in structured, validated, suite.rc files that concisely specify the properties of, and the relationships between, the various tasks managed by the suite. This section of the User Guide deals with the format and content of the suite.rc file, including task definition. Task implementation - what’s required of the real commands, scripts, or programs that do the processing that the tasks represent - is covered in 10; and task job submission - how tasks are submitted to run - is in 11.
A cylc suite definition directory contains:
A typical example:
Suite.rc files are an extended-INI format with section nesting.
Embedded template processor expressions may also be used in the file, to programatically generate the final suite definition seen by cylc. Currently the Jinja2 template processor is supported (http://jinja.pocoo.org/docs); see 9.6 for examples. In the future cylc may provide a plug-in interface to allow use of other template engines too.
The following defines legal suite.rc syntax:
Suites that embed Jinja2 code (see 9.6) must process to raw suite.rc syntax.
Cylc has native support for suite.rc include-files, which may help to organize large suites. Inclusion boundaries are completely arbitrary - you can think of include-files as chunks of the suite.rc file simply cut-and-pasted into another file. Include-files may be included multiple times in the same file, and even nested. Include-file paths can be specified portably relative to the suite definition directory, e.g.:
Editing Temporarily Inlined Suites Cylc’s native file inclusion mechanism supports optional inlined editing:
The suite will be split back into its constituent include-files when you exit the edit session. While editing, the inlined file becomes the official suite definition so that changes take effect whenever you save the file. See cylc prep edit --help for more information.
Include-Files via Jinja2 Jinja2 (9.6) also has template inclusion functionality.
Cylc comes with syntax files for a number of text editors:
Refer to comments at the top of each file to see how to use them.
Cylc suite.rc files consist of a suite title and description followed by configuration items grouped under several top level section headings:
Cylc suite.rc files are automatically validated against a specification that defines all legal entries, values, options, and defaults. This detects formatting errors, typographic errors, illegal items and illegal values prior to run time. Some values are complex strings that require further parsing by cylc to determine their correctness (this is also done during validation). All legal entries are documented in the Suite.rc Reference (A).
The validator reports the line numbers of detected errors. Here’s an example showing a section heading with a missing right bracket:
If the suite.rc file uses include-files cylc view will show an inlined copy of the suite with correct line numbers (you can also edit suites in a temporarily inlined state with cylc edit --inline).
Validation does not check the validity of chosen job submission methods.
The [scheduling] section of a suite.rc file defines the relationships between tasks in a suite - the information that allows cylc to determine when tasks are ready to run. The most important component of this is the suite dependency graph. Cylc graph notation makes clear textual graph representations that are very concise because sections of the graph that repeat at different hours of the day, say, only have to be defined once. Here’s an example with dependencies that vary depending on the particular cycle point:
Figure 20 shows the complete suite.rc listing alongside the suite graph. This is a complete, valid, runnable suite (it will use default task runtime properties such as script).
Multiline graph strings may contain:
Suite dependency graphs can be broken down into pairs in which the left side (which may be a single task or family, or several that are conditionally related) defines a trigger for the task or family on the right. For instance the “word graph” C triggers off B which triggers off A can be deconstructed into pairs C triggers off B and B triggers off A. In this section we use only the default trigger type, which is to trigger off the upstream task succeeding; see 9.3.4 for other available triggers.
In the case of cycling tasks, the triggers defined by a graph string are valid for cycle points matching the list of hours specified for the graph section. For example this graph,
implies that B triggers off A for cycle points in which the hour matches 00 or 12.
To define inter-cycle dependencies, attach an offset indicator to the left side of a pair:
This means B[time] triggers off A[time-PT12H] (12 hours before) for cycle points with hours matching 00 or 12. time is implicit because this keeps graphs clean and concise, given that the majority of tasks will typically depend only on others with the same cycle point. Cycle point offsets can only appear on the left of a pair, because a pairs define triggers for the right task at cycle point time. However, A => B[-PT6H], which is illegal, can be reformulated as a future trigger A[+PT6H] => B (see 9.3.4.11). It is also possible to combine multiple offsets within a cycle point offset e.g.
This means that B[Time] triggers off A[time-P1D-PT12H] (1 day and 12 hours before).
Triggers can be chained together. This graph:
is equivalent to this:
Each trigger in the graph must be unique but the same task can appear in multiple pairs or chains. Separately defined triggers for the same task have an AND relationship. So this:
is equivalent to this:
In summary, the branching tree structure of a dependency graph can be partitioned into lines (in the suite.rc graph string) of pairs or chains, in any way you like, with liberal use of internal white space and comments to make the graph structure as clear as possible.
Handling Long Graph Lines Long chains of dependencies can be split into pairs:
If you have very long task names, or long conditional trigger expressions (below) then you can use the suite.rc line continuation marker:
Note that a line continuation marker must be the final character on the line; it cannot be followed by trailing spaces or a comment.
A suite definition can contain multiple graph strings that are combined to generate the final graph.
One-off (Non-Cycling) Figure 21 shows a small suite of one-off non-cycling tasks; these all share a single cycle point (1) and don’t spawn successors (once they’re all finished the suite just exits). The integer 1 attached to each graph node is just an arbitrary label here.
Cycling Graphs For cycling tasks the graph section heading defines a sequence of cycle points for which the subsequent graph section is valid. Figure 22 shows a small suite of cycling tasks.
Advanced Cycling The hour-based section heading examples shown above (T00, T06, etc) are just one small part of cylc’s cycling syntax.
T06 stands for ”Repeat every 1 day starting at 06:00 after the initial cycle point”. Cylc allows you to start (or end) at any particular time, repeat at whatever frequency you like, and even optionally limit the number of repetitions.
For example, all these are valid in cylc:
Here is a quick summary of the syntax rules:
Date-time cycling information is made up of a starting date-time, an interval, and an optional limit.
The time is assumed to be in the local time zone unless you set [cylc]cycle point time zone or [cylc]UTC mode. The calendar is assumed to be the proleptic Gregorian calendar unless you set [scheduling]cycling mode.
The syntax for representations is based on the ISO 8601 date-time standard. This includes the representation of date-time, interval. What we define for cylc’s cycling syntax is our own optionally-heavily-condensed form of ISO 8601 recurrence syntax. The most common full form is: R[limit?]/[date-time]/[interval]. However, we allow omitting information that can be guessed from the context (rules below). This means that it can be written as:
For example, T00 is an example of [date-time], with an inferred 1 day period and no limit.
Where some or all date-time information is omitted, it is inferred to be relative to the initial date-time cycle point. For example, T00 by itself would mean the next occurrence of midnight that follows, or is, the initial cycle point. Entering +PT6H would mean 6 hours after the initial cycle point. Entering -P1D would mean 1 day before the initial cycle point. Entering no information for the date-time implies the initial cycle point date-time itself.
Where the interval is omitted and some (but not all) date-time information is omitted, it is inferred to be a single unit above the largest given specific date-time unit. For example, the largest given specific unit in T00 is hours, so the inferred interval is 1 day (daily), P1D.
Where the limit is omitted, unlimited cycling is assumed. This will be bounded by the final cycle point’s date-time if given.
Another supported form of ISO 8601 recurrence is: R[limit?]/[interval]/[date-time]. This form uses the date-time as the end of the cycling sequence rather than the start. For example, R3/P5D/20140430T06 means:
This kind of form can be used for specifying special behaviour near the end of the suite, at the final cycle point’s date-time. We can also represent this in cylc with a collapsed form:
So, for example, you can write:
How Multiple Graph Strings Combine For a cycling graph with multiple validity sections for different hours of the day, the different sections add to generate the complete graph. Different graph sections can overlap (i.e. the same hours may appear in multiple section headings) and the same tasks may appear in multiple sections, but individual dependencies should be unique across the entire graph. For example, the following graph defines a duplicate prerequisite for task C:
This does not affect scheduling, but for the sake of clarity and brevity the graph should be written like this:
Advanced Starting Up Dependencies that are only valid at the initial cycle point can be written using the R1 notation (e.g. as in 7.23.2. For example:
In the example above, R1 implies R1/20130808T00, so prep only runs once at that cycle point (the initial cycle point). At that cycle point, foo will have a dependence on prep - but not at subsequent cycle points.
However, it is possible to have a suite that has multiple effective initial cycles - for example, one starting at T00 and another starting at T12. What if they need to share an initial task?
Let’s suppose that we add the following section to the suite example above:
We’ll also say that there should be a starting dependence between prep and our new task baz - but we still want to have a single prep task, at a single cycle.
We can write this using a special case of the task[-interval] syntax - if the interval is null, this implies the task at the initial cycle point.
For example, we can write our suite like 23.
This neatly expresses what we want - a task running at the initial cycle point that has one-off dependencies with other task sets at different cycles.
A different kind of requirement is displayed in Figure 24. Usually, we want to specify additional tasks and dependencies at the initial cycle point. What if we want our first cycle point to be entirely special, with some tasks missing compared to subsequent cycle points?
In Figure 24, bar will not be run at the initial cycle point, but will still run at subsequent cycle points. [[[+PT6H/PT6H]]] means start at +PT6H (6 hours after the initial cycle point) and then repeat every PT6H (6 hours).
Some suites may have staggered start-up sequences where different tasks need running once but only at specific cycle points, potentially due to differing data sources at different cycle points with different possible initial cycle points. To allow this cylc provides a min( ) function that can be used as follows:
In this example the initial cycle point is 20100101T03, so the prep1 task will run once at 20100101T12 and the prep2 task will run once at 20100101T06 as these are the first cycle points after the initial cycle point in the respective min( ) entries.
Integer Cycling In addition to non-repeating and date-time cycling workflows, cylc can do integer cycling for repeating workflows that are not date-time based.
To construct an integer cycling suite, set [scheduling]cycling mode = integer, and specify integer values for the initial and (optional) final cycle points. The notation for intervals, offsets, and recurrences (sequences) is similar to the date-time cycling notation, except for the simple integer values.
The full integer recurrence expressions supported are:
But, as for date-time cycling, sequence start and end points can be omitted where suite initial and final cycle points can be assumed. Some examples:
The tutorial illustrates integer cycling in 7.23.3, and $CYLC_DIR/examples/satellite/ is a self-contained example of a realistic use for integer cycling. It simulates the processing of incoming satellite data: each new dataset arrives after a random (as far as the suite is concerned) interval, and is labeled by an arbitrary (as far as the suite is concerned) ID in the filename. A task called get_data at the top of the repeating workflow waits on the next dataset and, when it finds one, moves it to a cycle-point-specific shared workspace for processing by the downstream tasks. When get_data.1 finishes, get_data.2 triggers and begins waiting for the next dataset at the same time as the downstream tasks in cycle point 1 are processing the first one, and so on. In this way multiple datasets can be processed at once if they happen to come in quickly. A single shutdown task runs at the end of the final cycle to collate results. The suite graph is shown in Figure 25.
Trigger type, indicated by :type after the upstream task (or family) name, determines what kind of event results in the downstream task (or family) triggering.
Success Triggers The default, with no trigger type specified, is to trigger off the upstream task succeeding:
For consistency and completeness, however, the success trigger can be explicit:
Failure Triggers To trigger off the upstream task reporting failure:
Suicide triggers can be used to remove task B here if A does not fail, see 9.3.4.8.
Start Triggers To trigger off the upstream task starting to execute:
This can be used to trigger tasks that monitor other tasks once they (the target tasks) start executing. Consider a long-running forecast model, for instance, that generates a sequence of output files as it runs. A postprocessing task could be launched with a start trigger on the model (model:start => post) to process the model output as it becomes available. Note, however, that there are several alternative ways of handling this scenario: both tasks could be triggered at the same time (foo => model & post), but depending on external queue delays this could result in the monitoring task starting to execute first; or a different postprocessing task could be triggered off a message output for each data file (model:out1 => post1 etc.; see 9.3.4.5), but this may not be practical if the number of output files is large or if it is difficult to add cylc messaging calls to the model.
Finish Triggers To trigger off the upstream task succeeding or failing, i.e. finishing one way or the other:
Message Triggers Tasks can also trigger off custom output messages. These must be registered in the [runtime] section of the emitting task, and reported using the cylc message command in task scripting. The graph trigger notation refers to the item name of the registered output message. The example suite $CYLC_DIR/examples/message-triggers illustrates message triggering.
Job Submission Triggers It is also possible to trigger off a task submitting, or failing to submit:
A possible use case for submit-fail triggers: if a task goes into the submit-failed state, possibly after several job submission retries, another task that inherits the same runtime but sets a different job submission method and/or host could be triggered to, in effect, run the same job on a different platform.
Conditional Triggers AND operators (&) can appear on both sides of an arrow. They provide a concise alternative to defining multiple triggers separately:
OR operators (|) which result in true conditional triggers, can only appear on the left,2
Forecasting suites typically have simple conditional triggering requirements, but any valid conditional expression can be used, as shown in Figure 26 (conditional triggers are plotted with open arrow heads).
Suicide Triggers Suicide triggers take tasks out of the suite. This can be used for automated failure recovery. The suite.rc listing and accompanying graph in Figure 27 show how to define a chain of failure recovery tasks that trigger if they’re needed but otherwise remove themselves from the suite (you can run the AutoRecover.async example suite to see how this works). The dashed graph edges ending in solid dots indicate suicide triggers, and the open arrowheads indicate conditional triggers as usual. Suicide triggers are ignored by default in the graph view, unless you toggle them on with View -¿ Options -¿ Ignore Suicide Triggers.
Note that multiple suicide triggers combine in the same way as other triggers, so this:
is equivalent to this:
i.e. both foo and bar must succeed for baz to be taken out of the suite. If you really want a task to be taken out if any one of several events occurs then be careful to write it that way:
A word of warning on the meaning of “bare suicide triggers”. Consider the following suite:
Task bar has a suicide trigger but no normal prerequisites (a suicide trigger is not a task triggering prerequisite, it is a task removal prerequisite) so this is entirely equivalent to:
In other words both tasks will trigger immediately, at the same time, and then bar will be removed if foo succeeds.
If an active task proxy (currently in the submitted or running states) is removed from the suite by a suicide trigger, a warning will be logged.
Family Triggers Families defined by the namespace inheritance hierarchy ( 9.4) can be used in the graph trigger whole groups of tasks at the same time (e.g. forecast model ensembles and groups of tasks for processing different observation types at the same time) and for triggering downstream tasks off families as a whole. Higher level families, i.e. families of families, can also be used, and are reduced to the lowest level member tasks. Note that tasks can also trigger off individual family members if necessary.
To trigger an entire task family at once:
This is equivalent to:
To trigger other tasks off families we have to specify whether to triggering off all members starting, succeeding, failing, or finishing, or off any members (doing the same). Legal family triggers are thus:
Here’s how to trigger downstream processing after if one or more family members succeed, but only after all members have finished (succeeded or failed):
Writing Efficient Inter-Family Triggering While cylc allows writing dependencies between two families it is important to consider the number of dependencies this will generate. In the following example, each member of FAM2 has dependencies pointing at all the members of FAM1.
Expanding this out, you generate N ⋆ M dependencies, where N is the number of members of FAM1 and M is the number of members of FAM2. This can result in high memory use as the number of members of these families grows, potentially rendering the suite impractical for running on some systems.
You can greatly reduce the number of dependencies generated in these situations by putting dummy tasks in the graphing to represent the state of the family you want to trigger off. For example, if FAM2 should trigger off any member of FAM1 succeeding you can create a dummy task FAM1_succeed_any_marker and place a dependency on it as follows:
This graph generates only N + M dependencies, which takes significantly less memory and CPU to store and evaluate.
Inter-Cycle Triggers Typically most tasks in a suite will trigger off others in the same cycle point, but some may depend on others with other cycle points. This notably applies to warm-cycled forecast models, which depend on their own previous instances (see below); but other kinds of inter-cycle dependence are possible too.3 Here’s how to express this kind of relationship in cylc:
inter-cycle and trigger type (or message trigger) notation can be combined:
At suite start-up inter-cycle triggers refer to a previous cycle point that does not exist. This does not cause the dependent task to wait indefinitely, however, because cylc ignores triggers that reach back beyond the initial cycle point. That said, the presence of an inter-cycle trigger does normally imply that something special has to happen at start-up. If a model depends on its own previous instance for restart files, for instance, then an initial set of restart files has to be generated somehow or the first model task will presumably fail with missing input files. There are several ways to handle this in cylc using different kinds of one-off (non-cycling) tasks that run at suite start-up. They are illustrated in the Tutorial (7.23.1); to summarize here briefly:
R1, or R1/date-time tasks are the recommended way to specify unusual start up conditions. They allow you to specify a clean distinction between the dependencies of initial cycles and the dependencies of the subsequent cycles.
Initial tasks can be used for real model cold-start processes, whereby a warm-cycled model at any given cycle point can in principle have its inputs satisfied by a previous instance of itself, or by an initial task with (nominally) the same cycle point.
In effect, the R1 task masquerades as the previous-cycle-point trigger of its associated cycling task. At suite start-up initial tasks will trigger the first cycling tasks, and thereafter the inter-cycle trigger will take effect.
If a task has a dependency on another task in a different cycle point, the dependency can be written using the [offset] syntax such as [-PT12H] in foo[-PT12H] => foo. This means that foo at the current cycle point depends on a previous instance of foo at 12 hours before the current cycle point. Unlike the cycling section headings (e.g. [[[T00,T12]]]), dependencies assume that relative times are relative to the current cycle point, not the initial cycle point.
However, it can be useful to have specific dependencies on tasks at or near the initial cycle point. You can switch the context of the offset to be the initial cycle point by using the caret symbol: ^.
For example, you can write foo[^] to mean foo at the initial cycle point, and foo[^+PT6H] to mean foo 6 hours after the initial cycle point. Usually, this kind of dependency will only apply in a limited number of cycle points near the start of the suite, so you may want to write it in R1-based cycling sections. Here’s the example inter-cycle R1 suite from above again.
You can see there is a dependence on the initial R1 task prep for foo at the first T00 cycle point, and at the first T12 cycle point. Thereafter, foo just depends on its previous (12 hours ago) instance.
This is quite different to the non-preferred cold-start tasks way of working.
Special Cold-Start Tasks Special cold-start tasks are non-spawning tasks intended for use in general cycling sections, so they must be declared as special tasks. Dependence on them is retained throughout the suite run so they must be used in conditional triggers to avoid stalling the suite, like this:
Here, in the first cycle point model will trigger off the cold-start task cold_model. In subsequent cycle points model retains its dependence on cold_model, but the conditional expression allows it to trigger off the previous instance of itself instead. The reason for doing this is that it allows cold_model to be inserted manually later in the suite run to cold-start model again after problems that require skipping one or more cycles. If you do not need this mid-run cold start capability then it is simpler to start the suite with a normal task in an initial R1 recurrence, like this:
Special Sequential Tasks If a cycling task does not generate files required by its own successor, then successive instances can run in parallel if the opportunity arises. However, if such a task would interfere with its own siblings for internal reasons (e.g. use of a hardwired non cycle dependent temporary file or similar) then it can be forced to run sequentially. This can be done with explicit inter-cycle triggers in the graph:
or by declaring the task to be sequential:
The sequential declaration also results in each instance of foo triggering off its own predecessor, exactly as in the explicit version. The only difference is that implicit triggers will not appear in graph visualizations. The implicit version can also be considerably simpler when the task appears in multiple graph sections or in a non-uniform cycling sequence: this suite,
is equivalent to this one:
Future Triggers Cylc also supports inter-cycle triggering off tasks “in the future” (with respect to cycle point):
(Recall that A[t+PT6H] can run before B[t] because tasks in cylc have private cycle points). Future triggers present a problem at the suite shutdown rather than at start-up. Here, B at the final cycle point wants to trigger off an instance of A that will never exist because it is beyond the suite stop point. Consequently cylc prevents tasks from spawning successors that depend on other tasks beyond the stop point.
Clock Triggers In addition to depending on other tasks (and on external events - see 9.3.4.17) tasks can depend on the wall clock: specifically, they can trigger off a wall clock time expressed as an offset from their own cycle point:
Here, foo[2015-08-23T00] would trigger (other dependencies allowing) when the wall clock time reaches 2015-08-23T02. Clock-trigger offsets are normally positive, to trigger some time after the wall-clock time is equal to task cycle point.
Clock-triggers have no effect on scheduling if the suite is running sufficiently far behind the clock (e.g. after a delay, or because it is processing archived historical data) that the trigger times, which are relative to task cycle point, have already passed.
Clock-Expire Triggers Tasks can be configured to expire - i.e. to skip job submission and enter the expired state - if they are too far behind the wall clock when they become ready to run, and other tasks can trigger off this. As a possible use case, consider a cycling task that copies the latest of a set of files to overwrite the previous set: if the task is delayed by more than one cycle there may be no point in running it because the freshly copied files will just be overwritten immediately by the next task instance as the suite catches back up to real time operation. Clock-expire tasks are configured like clock-trigger tasks, with a date-time offset relative to cycle point (A.4.10.2). The offset should be positive to make the task expire if the wall-clock time has gone beyond the cycle point. Triggering off an expired task typically requires suicide triggers to remove the workflow that runs if the task has not expired. Here a task called copy expires, and its downstream workflow is skipped, if it is more than one day behind the wall-clock (see also examples/clock-expire):
External Triggers In addition to depending on other tasks (and on the wall clock - see 9.3.4.15) tasks can trigger off events reported by an external system. For example, an external process could detect incoming data on an ftp server, and then notify a suite containing a task to retrieve the new data for processing. This is an alternative to long-running tasks that poll for external events.
Note that cylc does not currently support triggering off “filesystem events” (e.g. inotify on Linux). However, external watcher processes can use filesystem events to detect triggering conditions, if that is appropriate, before notifying a suite with our general external event system.
External triggers are not normally needed in date-time cycling suites driven by real time data that comes in at regular intervals. In these cases a data retrieval task can be clock-triggered to submit at the expected data arrival time, so little time if any is wasted in polling. But in non-cycling or integer cycling suites (which can be used for arbitrarily timed repeating workflows such as for satellite data processing) external triggers may be preferred.
The external triggering process must call cylc ext-trigger with the name of the target suite, the message that identifies this type of event to the suite, and an ID that distinguishes this particular event instance from others (the name of the target task or its current cycle point is not required). The event ID is just an arbitary string to cylc, but it typically identifies the filename(s) of the latest dataset in some way. When the suite daemon receives the external event notification it will trigger the next instance of any task waiting on that trigger (whatever its cycle point) and then broadcast (see 12.19) the event ID to the cycle point of the triggered task as $CYLC_EXT_TRIGGER_ID. Downstream tasks with the same cycle point therefore know the new event ID too and can use it, if they need to, to identify the same new dataset. In this way a whole workflow can be associated with each new dataset, and multiple datasets can be processed in parallel if they happen to arrive in quick succession.
An externally-triggered task must register the event it waits on in the suite scheduling section:
Then, each time a new dataset arrives the external detection system should notify the suite like this:
where “sat-proc” is the suite name and “passX12334a” is the ID string for the new event. The suite passphrase must be installed on triggering account.
Note that only one task in a suite can trigger off a particular external message. Other tasks can trigger off the externally triggered task as required, of course.
$CYLC_DIR/examples/satellite/ext-triggers/suite.rc is a working example of a simulated satellite processing suite.
Warm-cycled forecast models generate restart files, e.g. model background fields, to initialize the next forecast. This kind of dependence requires an inter-cycle trigger:
Or if you don’t need an explicit cold-start task, this will do:
If your model is configured to write out additional restart files to allow one or more cycle points to be skipped in an emergency do not represent these potential dependencies in the suite graph as they should not be used under normal circumstances. For example, the following graph would result in task A erroneously triggering off A[T-24] as a matter of course, instead of off A[T-6], because A[T-24] will always be finished first:
The [runtime] section of a suite definition configures what to execute (and where and how to execute it) when each task is ready to run, in a multiple inheritance hierarchy of namespaces culminating in individual tasks. This allows all common configuration detail to be factored out and defined in one place.
Any namespace can configure any or all of the items defined in the Suite.rc Reference (A).
Namespaces that do not explicitly inherit from others automatically inherit from the root namespace (below).
Nested namespaces define task families that can be used in the graph as convenient shorthand for triggering all member tasks at once, or for triggering other tasks off all members at once - see 9.3.4.9. Nested namespaces can be progressively expanded and collapsed in the dependency graph viewer, and in the gcylc graph and text views. Only the first parent of each namespace (as for single-inheritance) is used for suite visualization purposes.
Namespace names may contain letters, digits, underscores, and hyphens.
Note that task names need not be hardwired into task implementations because task and suite identity can be extracted portably from the task execution environment supplied by the suite daemon (9.4.7) - then to rename a task you can just change its name in the suite definition.
The root namespace, at the base of the inheritance hierarchy, provides default configuration for all tasks in the suite. Most root items are unset by default, but some have default values sufficient to allow test suites to be defined by dependency graph alone. The script item, for example, defaults to code that prints a message then sleeps for between 1 and 15 seconds and exits. Default values are documented with each item in A. You can override the defaults or provide your own defaults by explicitly configuring the root namespace.
If a namespace section heading is a comma-separated list of names then the subsequent configuration applies to each list member. Particular tasks can be singled out at run time using the $CYLC_TASK_NAME variable.
As an example, consider a suite containing an ensemble of closely related tasks that each invokes the same script but with a unique argument that identifies the calling task name:
For large ensembles Jinja2 template processing can be used to automatically generate the member names and associated dependencies (see 9.6).
The following listing of the inherit.single.one example suite illustrates basic runtime inheritance with single parents.
If a namespace inherits from multiple parents the linear order of precedence (which namespace overrides which) is determined by the so-called C3 algorithm used to find the linear method resolution order for class hierarchies in Python and several other object oriented programming languages. The result of this should be fairly obvious for typical use of multiple inheritance in cylc suites, but for detailed documentation of how the algorithm works refer to the official Python documentation here: http://www.python.org/download/releases/2.3/mro/.
The inherit.multi.one example suite, listed here, makes use of multiple inheritance:
cylc get-suite-config provides an easy way to check the result of inheritance in a suite. You can extract specific items, e.g.:
or use the --sparse option to print entire namespaces without obscuring the result with the dense runtime structure obtained from the root namespace:
Suite Visualization And Multiple Inheritance The first parent inherited by a namespace is also used as the collapsible family group when visualizing the suite. If this is not what you want, you can demote the first parent for visualization purposes, without affecting the order of inheritance of runtime properties:
The linear precedence order of ancestors is computed for each namespace using the C3 algorithm. Then any runtime items that are explicitly configured in the suite definition are “inherited” up the linearized hierarchy for each task, starting at the root namespace: if a particular item is defined at multiple levels in the hierarchy, the level nearest the final task namespace takes precedence. Finally, root namespace defaults are applied for every item that has not been configured in the inheritance process (this is more efficient than carrying the full dense namespace structure through from root from the beginning).
The task execution environment contains suite and task identity variables provided by the suite daemon, and user-defined environment variables. The environment is explicitly exported (by the task job script) prior to executing the task script (see 11).
Suite and task identity are exported first, so that user-defined variables can refer to them. Order of definition is preserved throughout so that variable assignment expressions can safely refer to previously defined variables.
Additionally, access to cylc itself is configured prior to the user-defined environment, so that variable assignment expressions can make use of cylc utility commands:
User Environment Variables A task’s user-defined environment results from its inherited [[[environment]]] sections:
This results in a task foo with SHAPE=circle, COLOR=blue, and TEXTURE=rough in its environment.
Overriding Environment Variables When you override inherited namespace items the original parent item definition is replaced by the new definition. This applies to all items including those in the environment sub-sections which, strictly speaking, are not “environment variables” until they are written, post inheritance processing, to the task job script that executes the associated task. Consequently, if you override an environment variable you cannot also access the original parent value:
The compressed variant of this, COLOR = dark-$COLOR, is also in error for the same reason. To achieve the desired result you must use a different name for the parent variable:
Suite And Task Identity Variables The task identity variables provided to tasks by the suite daemon include:
And the suite identity variables are:
Some of these variables are also used by cylc task messaging commands in order to target the right task proxy object in the right suite.
Suite Share Directories A suite share directory is created automatically under the suite run directory as a share space for tasks. The location is available to tasks as $CYLC_SUITE_SHARE_DIR. In a cycling suite, output files are typically held in cycle point sub-directories of the suite share directory.
The top level share and work directory (below) location can be changed (e.g. to a large data area) by a global config setting (see B.9.1.2).
Task Work Directories Task job scripts are executed from within work directories created automatically under the suite run directory. A task can get its own work directory from $CYLC_TASK_WORK_DIR (or simply $PWD if it does not cd elsewhere at runtime). By default the location contains task name and cycle point, to provide a unique workspace for every instance of every task. This can be overridden in the suite definition, however, to get several tasks to share the same work directory (see A.5.1.13).
The top level work and share directory (above) location can be changed (e.g. to a large data area) by a global config setting (see B.9.1.2).
Other Cylc-Defined Environment Variables Initial and final cycle points, if supplied via the suite.rc file or the command line, are passed to task execution environments as:
Tasks can use these to determine whether or not they are running in the first or final cycles. Note however that R1 graph sections are now the preferred way to get different behaviour at suite start-up or shutdown.
Environment Variable Evaluation Variables in the task execution environment are not evaluated in the shell in which the suite is running prior to submitting the task. They are written in unevaluated form to the job script that is submitted by cylc to run the task (11.1) and are therefore evaluated when the task begins executing under the task owner account on the task host. Thus $HOME, for instance, evaluates at run time to the home directory of task owner on the task host.
Tasks can use $CYLC_SUITE_DEF_PATH to access suite files on the task host, and the suite bin directory is automatically added $PATH. If a remote suite definition directory is not specified the local (suite host) path will be assumed with the local home directory, if present, swapped for literal $HOME for evaluation on the task host.
If a task declares an owner other than the suite owner and/or a host other than the suite host, cylc will use non-interactive ssh to execute the task on the owner@host account by the configured job submission method,
For this to work,
To learn how to give remote tasks access to cylc, see 12.7.
Tasks running on the suite host under another user account are treated as remote tasks.
Remote hosting, like all namespace settings, can be declared globally in the root namespace, or per family, or for individual tasks.
Dynamic Host Selection Instead of hardwiring host names into the suite definition you can specify a shell command that prints a hostname, or an environment variable that holds a hostname, as the value of the host config item. See A.5.1.19.1.
Remote Task Log Directories Task stdout and stderr streams are written to log files in a suite-specific sub-directory of the suite run directory, as explained in 11.3. For remote tasks the same directory is used, but on the task host. Remote task log directories, like local ones, are created on the fly, if necessary, during job submission.
The visualization section of a suite definition is used to configure suite graphing, principally graph node (task) and edge (dependency arrow) style attributes. Tasks can be grouped for the purpose of applying common style attributes. See A for details.
Nested families from the runtime inheritance hierarchy can be expanded and collapsed in suite graphs and the gcylc graph view. All families are displayed in the collapsed state at first, unless [visualization]collapsed families is used to single out specific families for initial collapsing.
In the gcylc graph view, nodes outside of the main graph (such as the members of collapsed families) are plotted as rectangular nodes to the right if they are doing anything interesting (submitted, running, failed).
Figure 28 illustrates successive expansion of nested task families in the namespaces example suite.
Cylc has built in support for the Jinja2 template processor in suite definitions. Jinja2 variables, mathematical expressions, loop control structures, conditional logic, etc., are automatically processed to generate the final suite definition seen by cylc.
The need for Jinja2 processing must be declared with a hash-bang comment as the first line of the suite.rc file:
Potential uses for this include automatic generation of repeated groups of similar tasks and dependencies, and inclusion or exclusion of entire suite sections according to the value of a single flag. Consider a large complicated operational suite and several related parallel test suites with slightly different task content and structure (the parallel suites, for instance, might take certain large input files from the operation or the archive rather than downloading them again) - these can now be maintained as a single master suite definition that reconfigures itself according to the value of a flag variable indicating the intended use.
Template processing is the first thing done on parsing a suite definition so Jinja2 expressions can appear anywhere in the file (inside strings and namespace headings, for example).
Jinja2 is well documented at http://jinja.pocoo.org/docs, so here we just provide an example suite that uses it. The meaning of the embedded Jinja2 code should be reasonably self-evident to anyone familiar with standard programming techniques.
The jinja2.ensemble example, graphed in Figure 29, shows an ensemble of similar tasks generated using Jinja2:
Here is the generated suite definition, after Jinja2 processing:
And finally, the jinja2.cities example uses variables, includes or excludes special cleanup tasks according to the value of a logical flag, and it automatically generates all dependencies and family relationships for a group of tasks that is repeated for each city in the suite. To add a new city and associated tasks and dependencies simply add the city name to list at the top of the file. The suite is graphed, with the New York City task family expanded, in Figure 30.
This functionality is not provided by Jinja2 by default, but cylc automatically imports the user environment to the template in a dictionary structure called environ. A usage example:
This example is emphasizes that the environment is read on the suite host at the time the suite definition is parsed - it is not, for instance, read at task run time on the task host.
Jinja2 variable values can be modified by “filters”, using pipe notation. For example, the built-in trim filter strips leading and trailing white space from a string:
(See official Jinja2 documentation for available built-in filters.)
Cylc also supports custom Jinja2 filters. A custom filter is a single Python function in a source file with the same name as the function (plus “.py” extension) and stored in one of the following locations:
In the filter function argument list, the first argument is the variable value to be “filtered”, and subsequent arguments can be whatever is needed. Currently there is one custom filter called “pad” in the central cylc Jinja2 filter directory, for padding string values to some constant length with a fill character - useful for generating task names and related values in ensemble suites:
Associative arrays (dicts in Python) can be very useful. Here’s an example, from
$CYLC_DIR/examples/jinja2/dict:
Here’s the result:
The values of Jinja2 variables can be passed in from the cylc command line rather than hardwired in the suite
definition. Here’s an example, from
$CYLC_DIR/examples/jinja2/defaults:
Here’s the result:
Note also that cylc view --set FIRST_TASKbob –jinja2 SUITE= will show the suite with the Jinja2 variables as set.
Warning: suites started with template variables set on the command line do not currently restart with the same settings - you have to set them again on the cylc restart command line.
It is sometimes convenient to omit certain tasks from the suite at runtime without actually deleting their definitions from the suite.
Defining [runtime] properties for tasks that do not appear in the suite graph results in verbose-mode validation warnings that the tasks are disabled. They cannot be used because the suite graph is what defines their dependencies and valid cycle points. Nevertheless, it is legal to leave these orphaned runtime sections in the suite definition because it allows you to temporarily remove tasks from the suite by simply commenting them out of the graph.
To omit a task from the suite at runtime but still leave it fully defined and available for use (by insertion or cylc submit) use one or both of [scheduling][[special task]] lists, include at start-up or exclude at start-up (documented in A.4.10.8 and A.4.10.7). Then the graph still defines the validity of the tasks and their dependencies, but they are not actually loaded into the suite at start-up. Other tasks that depend on the omitted ones, if any, will have to wait on their insertion at a later time or otherwise be triggered manually.
Finally, with Jinja2 (9.6) you can radically alter suite structure by including or excluding tasks from the [scheduling] and [runtime] sections according to the value of a single logical flag defined at the top of the suite.
A naked dummy task appears in the suite graph but has no explicit runtime configuration section. Such tasks automatically inherit the default “dummy task” configuration from the root namespace. This is very useful because it allows functional suites to be mocked up quickly for test and demonstration purposes by simply defining the graph. It is somewhat dangerous, however, because there is no way to distinguish an intentional naked dummy task from one generated by typographic error: misspelling a task name in the graph results in a new naked dummy task replacing the intended task in the affected trigger expression; and misspelling a task name in a runtime section heading results in the intended task becoming a dummy task itself (by divorcing it from its intended runtime config section).
To avoid this problem any dummy task used in a real suite should not be naked - i.e. it should have an explicit entry in under the runtime section of the suite definition, even if the section is empty. This results in exactly the same dummy task behaviour, via implicit inheritance from root, but it allows use of cylc validate --strict to catch errors in task names by failing the suite if any naked dummy tasks are detected.
Existing scripts or executables can be used as cylc tasks without any modification so long as they return standard exit status on success or failure, and do not internally spawn detaching processes (see 10.4).
Simple tasks can be entirely implemented within the suite.rc file, as the task script items can be multi-line strings.
Tasks should abort with non-zero exit status if a fatal error occurs (this is just good coding practice anyway). This allows cylc’s task job scripts to automatically trap errors and send cylc task message "failed" back to the suite. The shell set -e option can be used in lieu of explicit error checks for every command:
General (non-output) messages can also be sent to report progress, warnings, and so on, e.g.:
Explanatory messages can be sent before aborting on error:
Or equivalently, with different syntax:
But not this:
If critical errors are not reported in this way task failures will still be detected and logged by the suite daemon, but you may have to examine task logs to determine what the problem was.
If a task script starts background sub-processes and does not wait on them, or internally submits jobs to a batch scheduler and then exits immediately, the detached processes will not be visible to cylc and the task will appear to finish when the top-level script finishes. You will need to modify scripts like this to make them execute all sub-processes in the foreground (or use the shell wait command to wait on them before exiting) and to prevent job submission commands from returning before the job completes (e.g. llsubmit -s for Loadleveler, qsub -sync yes for Sun Grid Engine, and qsub -W block=true for PBS).
If this is not possible - perhaps you don’t have control over the script or can’t work out how to fix it - one alternative approach is to use another task to repeatedly poll for the results of the detached processes:
For the requirements a command, script, or program, must fulfill in order to function as a cylc task, see 10. This section explains how tasks are submitted by the suite daemon when they are ready to run, and how to define new task job submission methods.
When a task is ready cylc generates a task job script that encapsulates the runtime settings (environment and scripting) defined for it in the suite.rc file. The job script is submitted to run by the job submission method chosen for the task. Different tasks can have different job submission methods. Like other runtime properties, you can set a suite default job submission method and override it for specific tasks or families:
As shown in the Tutorial (7.11) job scripts are saved to the suite run directory; the commands used to submit them are printed to stdout by cylc in debug mode; and they can be printed with the cylc log command or new ones generated and printed with the cylc jobscript command. Take a look at one to see exactly how cylc wraps and runs your tasks.
Cylc supports a number of commonly used job submission methods. See 11.7 for how to add new job submission methods.
Runs tasks directly in a background shell.
Submits tasks to the rudimentary Unix at scheduler. The atd daemon must be running.
Submits tasks to loadleveler by the llsubmit command. Loadleveler directives can be provided in the suite.rc file:
These are written to the top of the task job script like this:
If restart=yes is specified as a directive for loadleveler, the job will automatically trap SIGUSR1, which loadleveler may use to preempt the job. On trapping SIGUSR1, the job will inform the suite that it has been vacated by loadleveler. This will put it back to the submitted state, until it starts running again.
Submits tasks to IBM Platform LSF by the bsub command. LSF directives can be provided in the suite.rc file:
These are written to the top of the task job script like this:
Submits tasks to PBS (or Torque) by the qsub command. PBS directives can be provided in the suite.rc file:
These are written to the top of the task job script like this:
Submits tasks to the Moab workload manager by the msub command. Moab directives can be provided in the suite.rc file; the syntax is very similar to PBS:
These are written to the top of the task job script like this:
(Moab understands #PBS directives).
Submits tasks to Sun/Oracle Grid Engine by the qsub command. SGE directives can be provided in the suite.rc file:
These are written to the top of the task job script like this:
Submits tasks to Simple Linux Utility for Resource Management by the sbatch command. SLURM directives can be provided in the suite.rc file (note that since not all SLURM commands have a short form, cylc requires the long form directives):
These are written to the top of the task job script like this:
For job submission methods that use job file directives (PBS, Loadlevler, etc.) default directives are provided to set the job name and stdout and stderr file paths.
As shown in the example above, to get a naked option flag such as -V in PBS or -cwd in SGE you must give a null string as the directive value in the suite.rc file.
When a task is ready to run cylc generates a filename root to be used for the task job script and log files. The filename containing the task name, cycle point, and a submit number that increments if the same task is re-triggered multiple times:
How the stdout and stderr streams are directed into these files depends on the job submission method. The background method just uses appropriate output redirection on the command line, as shown above. The loadleveler method writes appropriate directives to the job script that is submitted to loadleveler.
Cylc obviously has no control over the stdout and stderr output from tasks that do their own internal output management (e.g. tasks that submit internal jobs and direct the associated output to other files). For less internally complex tasks, however, the files referred to here will be complete task job logs.
Some job submission methods, such as pbs, redirect a job’s stdout and stderr streams to a separate cache area while the job is running. The contents are only copied to the normal locations when the job completes. This means that cylc log or the gcylc GUI will be unable to find the job’s stdout and stderr streams while the job is running. Some sites with these job submission methods are known to provide commands for viewing and/or tail-follow a job’s stdout and stderr streams that are redirected to these cache areas. If this is the case at your site, you can configure cylc to make use of the provided commands by adding some settings to the global site/user config. E.g.:
To change the form of the actual command used to submit a job you do not need to define a new job submission method; just override the command template in the relevant job submission sections of your suite.rc file:
As explained in A the template’s %(job)s will be substituted by the job file path.
For supported job submission methods, one-way polling can be used to determine actual job status: the suite daemon executes a process on the task host, by non-interactive ssh, to interrogate the batch queueing system there, and to read a status file that is automatically generated by the task job script as it runs.
Polling may be required to update the suite state correctly after unusual events such as a machine being rebooted with tasks running on it, or network problems that prevent task messages from getting back to the suite host.
Tasks can be polled on demand by right-clicking on them in gcylc or using the cylc poll command.
Tasks are polled automatically, once, if they timeout during job submission or execution (see A.5.1.20 for how to configure timeouts).
Any tasks recorded in the submitted or running states at suite restart are automatically polled to determine what happened to them while the suite was down.
Regular polling can also be configured as a health check on tasks submitted to hosts that are known to be flaky, or as the sole method of determining task status on hosts that do not allow task messages to be routed back to the suite host.
To use polling instead of task-to-suite messaging set task communication method = poll in cylc site and user global config (see B.9.1.3). The default polling intervals can be overridden for all suites there too (see B.1.8 and B.1.7), or in specific suite definitions (in which case polling will be done regardless of the task communication method configured for the host; see A.5.1.11 and A.5.1.12).
Note that regular polling is not as efficient as task messaging in updating task status, and it should be used sparingly in large suites.
For supported job submission methods, the suite daemon can execute a process on the task host, by non-interactive ssh, to kill a submitted or running job according to its job submission method.
Tasks can be killed on demand by right-clicking on them in gcylc or using the cylc kill command.
Defining a new job submission method requires a little Python programming. You can use one of the built-in methods as an example, and read the documentation in the header of the cylc.batch_sys_manager module.
The following user-defined job submission class, called qsub, overrides the built-in pbs class to change the directive prefix from #PBS to #QSUB:
To check that this works correctly save the new source file to qsub.py in one of the allowed locations (see just below), use it in a suite definition:
and generate a job script to see the resulting directives:
You new job submission class code should be saved to a file with the same name as the class (plus “.py” extension). It can reside in any of the following locations, depending on how generally useful the new method is and whether or not you have write-access to the cylc source tree:
This chapter currently features a diverse collection of topics related to running suites. Please also see the Tutorial (7) and command documentation (D), and experiment with plenty of examples.
There are three ways to start a suite running: cold start and warm start, which start from scratch; and restart, which loads a prior suite state. If a suite does not contain any special cold-start tasks there is no difference between cold and warm start, except that a warm start should start from a point beyond the original initial cycle point of the suite. Special cold-start tasks are only needed if you want to be able to cold-start individual tasks again in the middle of a suite run - see 9.3.4.12.
Once a suite is up and running it is typically a restart that is needed most often (but see also cylc reload). Be aware that cold and warm starts wipe out any prior suite state, which prevents returning to a restart if you decide that’s what you really intended.
A cold start is the primary way to start a suite run from scratch:
The initial cycle point may be specified on the command line or in the suite.rc file. The scheduler starts by loading the first instance of each task at the suite initial cycle point, or at the next valid point for the task, including any special cold-start tasks (see the note on cold-start tasks at the beginning of this section).
A restart starts a suite run from the state recorded at the end of a previous run. This allows restarting a suite that was shut down or killed, without rerunning tasks that were already completed, or which were already submitted or running when the suite went down.
For a restart, the scheduler starts by loading each task in its recorded state. The most recent state is loaded by default, but earlier state files can be specified on the command line. Any tasks recorded as ‘submitted’ or ‘running’ will be polled automatically to determine what happened to them while the suite was down.
A warm start runs a suite from scratch like a cold start, but from a given cycle point that is later than the suite’s initial cycle point. All tasks from the given cycle point will run. It can be considered an inferior alternative to a restart because it may result in some tasks rerunning. A warm start may be required if a restart is not possible because the suite state files were accidentally deleted (for instance). The warm start cycle point must be given on the command line:
The original suite initial cycle point is preserved, but all tasks and dependencies before the given start cycle point are ignored.
The scheduler starts by loading a first instance of each task at the warm start cycle point, or at the next valid point for the task. Special cold-start tasks are loaded in the ‘succeeded’ state (this is merely a device to satisfy any initial dependence on the cold-start tasks, which are assumed not to be needed for a warm start; see the note on cold-start tasks at the beginning of this section). R1-type tasks behave exactly the same as other tasks - if their cycle point is at or later than the given start cycle point, they will run; if not, they will be ignored.
Cylc has three ways of tracking the progress of tasks, configured per task host in the site and user global config files (6). All three methods can be used on different task hosts within the same suite if necessary.
The Pyro communication method is the default because it is the most direct and efficient; the ssh method inserts an extra step in the process (command re-invocation on the suite host); and task polling is the least efficient because results are checked at predetermined intervals, not when task events actually occur.
Be careful to avoid spamming task hosts with polling commands. Each poll opens (and then closes) a new ssh connection.
Polling intervals are configurable under [runtime] because they should may depend on the expected execution time. For instance, a task that typically takes an hour to run might be polled every 10 minutes initially, and then every minute toward the end of its run. Interval values are used in turn until the last value, which is used repeatedly until finished:
A list of intervals with optional multipliers can be used for both submission and execution polling, although a single value is probably sufficient for submission polling. If these items are not configured default values from site and user global config will be used for the polling task communication method; polling is not done by default under the other task communications methods (but it can still be used if you like).
Polling is also done automatically once on job submission and execution timeouts, to see if the timed-out task has failed or not; and on suite restarts, to see what happened to any tasks that were orphaned when the suite went down.
If Pyro and ssh ports are blocked but you don’t want to use polling from the suite host,
Here are the default site and user global config items relevant to task state tracking (see these with cylc get-site-config):
User-invoked commands that connect to running suites can also choose between direct communication across network sockets (Pyro) and re-invocation of commands on the suite host using non-interactive ssh (there is a --use-ssh command option for this purpose).
The gcylc GUI requires direct Pyro connections to its target suite. If that is not possible, run gcylc on the suite host.
Suite daemons listen on dedicated network ports for incoming client requests - task messages and user-invoked commands (CLI or GUI). The cylc scan command reveals which suites are running on scanned hosts, and what ports they are listening on.
Client programs have to authenticate with the target suite daemon before issuing commands or requesting information. Cylc has two authentication levels: full read and control via a suite-specific passphrase (see 12.6.1); and configurable free “public” access (see 12.6.2).
A file called passphrase is generated in the suite definition directory at registration time, containing a single line of random text. The passphrase is loaded by the suite daemon at start-up and used to authenticate client connections. Suite passphrases are used in an encrypted challenge-response scheme; they are never sent raw over the network.
Clients on the suite host account automatically pick up the passphrase from the suite definition directory. On submission of the first task job on a host, the suite host will attempt to install the passphrase to the run directory to enable task jobs to connect to the suite. On other accounts that do not share the same HOME directory with the suite host account and do not have non-interactive SSH access to the suite host account, it is possible to install the passphrase manually. Allowed passphrase locations are:
Note: A client will attempt to install the passphrase automatically via non-interactive SSH to the first location listed above.
Possession of a suite passphrase gives full read and control access to the suite. Without the passphrase the amount of information revealed by a suite daemon is determined by the public access privilege level set in global site/user config (B.15) and optionally overidden in suites (A.3):
The default public access level is state-totals. The cylc scan command can print descriptions and task state totals in addition to basic suite identity, if you have the right passphrases or if the requested information is revealed publicly.
Running tasks need access to cylc via $PATH, principally for the task messaging commands. To allow this, the first thing a task job script does is set $CYLC_VERSION to the cylc version number of the running suite. If you need to run several suites at once under different incompatible versions of cylc, check that your site is using the cylc version wrapper (see INSTALL and admin/cylc-wrapper in a cylc installation) then set $CYLC_VERSION to the desired version. In the case of developers wishing to run their own copy of cylc rather than a centrally installed one, set $CYLC_HOME to point to your cylc copy.
Access to the cylc executable for different hosts can be configured using the site and user global configuration files. If the environment for running the cylc executable is only set up correctly in a login shell for a given host, you can set the [hosts]\textrightarrow HOST \textrightarrow cylc executable setting for the relevant host to True. (This is the default behaviour.) If the environment is already correct without the login shell, but the cylc executable is not in $PATH, then the [hosts]\textrightarrow HOST \textrightarrow cylc executable setting can be used to specify the path to the cylc executable.
The site/user global config item global init-script can also be used if necessary (see B.9.1.8).
A restarted suite (see cylc restart --help) is initialized from a previous recorded suite state dump so that it can carry on from wherever it got to before being shut down or killed.
Tasks that were recorded in the submitted or running states are now automatically polled on restart, to see if they are still submitted (e.g. waiting in a PBS batch queue or similar), still running, or if they finished (succeeded or failed) while the suite was down.
Tasks recorded in the failed state at shutdown are not automatically resubmitted on restarting the suite, in case the underlying problem has not been addressed yet.
As a suite runs its task proxies may pass through the following states:
Note that greyed-out “base graph nodes” in the gcylc graph view do not represent task states; they are displayed to fill out the graph structure where corresponding task proxies do not currently exist in the live task pool.
For manual task state reset purposes ready is a pseudo-state that means waiting with all prerequisites satisfied.
Connecting to a running suite requires knowing the network port it is listening on, and the suite passphrase to authenticate with once a connection is made to the port.
Suites write their port number to $HOME/.cylc/ports/<SUITE> at start-up, and suite-connecting commands read this file to get the number.4 An exception to this is the messaging commands called by tasks. Running tasks know the port number from the execution environment provided by the suite (via the task job script).
So, to connect to a suite running on another account you must install the suite passphrase (12.6.1), and configure non-interactive ssh so that the port number can be retrieved from the remote port file. Then use the --user and --host command options to connect:
Alternatively, you can determine suite port numbers using cylc scan, and use them explicitly on the command line:
Possession of a suite passphrase gives full control over the suite, and ssh access to the port file also implies full access to the suite host account, so it is recommended that this only be used to interact with your own suites running on other hosts. We plan to implement finer-grained authentication in the future to allow suite owners to grant read-only access to others.
A connection timeout can be set in site and user global config files (see 6) so that messaging commands cannot hang indefinitely if the suite is not responding (this can be caused by suspending a suite with Ctrl-Z) thereby preventing the task from completing. The same can be done on the command line for other suite-connecting user commands, with the --pyro-timeout option.
Runahead limiting prevents the fastest tasks in a suite from getting too far ahead of the slowest ones. Newly spawned tasks are released to the task pool only when they fall below the runahead limit. A low runhead limit can prevent cylc from interleaving cycles, but it will not stall a suite unless it fails to extend out past a future trigger (see 9.3.4.11). A high runahead limit may allow fast tasks that are not constrained by dependencies or clock-triggers to spawn far ahead of the pack, which could have performance implications for the suite daemon when running very large suites. Succeeded and failed tasks are ignored when computing the runahead limit.
The preferred runahead limiting mechanism restricts the number of consecutive active cycle points. The default value is three active cycle points; see A.4.8. Alternatively the interval between the slowest and fastest tasks can be specified as hard limit; see A.4.7.
Large suites can potentially overwhelm task hosts by submitting too many tasks at once. You can prevent this with internal queues, which limit the number of tasks that can be active (submitted or running) at the some time.
A queue is defined by a name; a limit, which is the maximum number of active tasks allowed for the queue; and a list of members, assigned by task or family name.
Queue configuration is done under the [scheduling] section of the suite.rc file (like dependencies, internal queues constrain when a task runs).
By default every task is assigned to the default queue, which by default has a zero limit (interpreted by cylc as no limit). To use a single queue for the whole suite just set the default queue limit:
To use additional queues just name each one, set their limits, and assign members:
Any tasks not assigned to a particular queue will remain in the default queue. The queues example suite illustrates how queues work by running two task trees side by side (as seen in the graph GUI) each limited to 2 and 3 tasks respectively:
See also A.5.1.10 in the Suite.rc Reference.
Tasks can be configured with a list of “retry delay” periods, as ISO 8601 durations, such that if a task fails it will go into a temporary retrying state and then automatically resubmit itself after the next specified delay period expires. A usage example is shown in the suite listed below under 12.15.
See also A.2.9, A.5.1.20 and A.5.1.21 in the Suite.rc Reference.
Cylc can call nominated event handlers when certain suite or task events occur. This is intended to facilitate centralized alerting and automated handling of critical events. Event handlers can be used to send a message, call a pager, and so on; or intervene in the operation of their own suite using cylc commands.
To send an email, you can use the built-in setting [[[events]]]mail events to specify a list of events for which notifications should be sent. E.g. to send an email on (submission) failed and retry:
By default, the emails will be sent to the current user with:
These can be configured using the settings:
Event handler commands can be located in the suite bin/ directory, otherwise it is up to you to ensure their location is in $PATH (in the shell in which cylc runs, on the suite host). The commands should require very little resource to run and should return quickly. (Each event handler is invoked by a child process in a finite process pool that is also used to submit, poll and kill jobs. The child process will wait for the event handler to complete before moving on to the next item in the queue. If the process pool is saturated with long running event handlers, the suite will appear to hang.)
Task event handlers can be specified using the [[[event hooks]]]<event> handler settings, where <event> is one of:
The value of each setting should be a list of command lines or command line templates (see below).
Alternatively, task event handlers can be specified using the [[[events]]]handlers and the [[[events]]]handler events settings, where the former is a list of command lines or command line templates (see below) and the latter is a list of events for which these commands should be invoked.
A command line template may have any or all of these patterns which will be substituted with actual values:
Otherwise, the command line will be called with the following command line arguments:
For an explanation of the substitution syntax, see String Formatting Operations in the Python documentation.
The retry event occurs if a task fails and has any remaining retries configured (see 12.14). The event handler will be called as soon as the task fails, not after the retry delay period when it is resubmitted.
Note that event handlers are called by cylc itself, not by the running tasks so if you wish to pass them additional information via the environment you must use [cylc] →[[environment]], not task runtime environments.
The following 2 suite.rc snippets are examples on how to specify event handlers using the alternate methods:
Note: The handler command is called like this:
The cylc reload command reloads the suite definition at run time. This allows: (a) changing task config items such as script or environment; (b) adding tasks to, or removing them from, the suite definition, at run time - without shutting the suite down and restarting it. (It is easy to shut down and restart cylc suites, but reloading may be useful if you don’t want to wait for long-running tasks to finish first).
Note that defined tasks can be already be added to or removed from a running suite with the cylc insert and cylc remove commands; the reload command allows addition and removal of task definitions. If a new task is definition is added (and used in the graph) you will still need to manually insert an instance of it (with a particular cycle point) into the running suite. If a task definition (and its use in the graph) is deleted, existing task proxies of the of the deleted type will run their course after the reload but new instances will not be spawned. Changes to a task definition will only take effect when the next task instance is spawned (existing instances will not be affected).
Some HPC facilities allow job preemption: the resource manager can kill or suspend running low priority jobs in order to make way for high priority jobs. The preempted jobs may then be automatically restarted by the resource manager, from the same point (if suspended) or requeued to run again from the start (if killed). If a running cylc task gets suspended or hard-killed (kill -9 <PID> is not a trappable signal so cylc cannot detect task failure in this case) and then later restarted, it will just appear to cylc as if it takes longer than normal to run. If the job is soft-killed the signal will be trapped by the task job script and a failure message sent, resulting in cylc putting the task into the failed state. When the preempted task restarts and sends its started message cylc would normally treat this as an error condition (a dead task is not supposed to be sending messages) - a warning will be logged and the task will remain in the failed state. However, if you know that preemption is possible on your system you can tell cylc that affected tasks should be resurrected from the dead, to carry on as normal if progress messages start coming in again after a failure:
To test this in any suite, manually kill a running task then, after cylc registers the task failed, resubmit the killed job manually by cutting-and-pasting the original job submission command from the suite stdout stream.
Any task proxy currently present in the suite can be manually triggered at any time using the cylc trigger command, or from the right-click task menu in gcylc. If the task belongs to a limited internal queue (see 12.13), this will queue it; if not, or if it is already queued, it will submit immediately.
With cylc trigger --edit (also in the gcylc right-click task menu) you can edit the generated task job script to make one-off changes before the task submits.
The cylc broadcast command overrides [runtime] settings in a running suite. This can be used to communicate information to downstream tasks by broadcasting environment variables (communication of information from one task to another normally takes place via the filesystem, i.e. the input/output file relationships embodied in inter-task dependencies). Variables (and any other runtime settings) may be broadcast to all subsequent tasks, or targeted specifically at a specific task, all subsequent tasks with a given name, or all tasks with a given cycle point; see broadcast command help for details.
Broadcast settings targeted at a specific task ID or cycle point expire and are forgotten as the suite moves on. Un-targeted variables and those targetted at a task name persist throughout the suite run, even across restarts, unless manually cleared using the broadcast command - and so should be used sparingly.
When a suite is started with the cylc run command (cold or warm start) the cycle point at which it starts can be given on the command line or hardwired into the suite.rc file:
or,
An initial cycle given on the command line will override one in the suite.rc file.
In the case of a cold start only the initial cycle point is passed through to task execution environments as $CYLC_SUITE_INITIAL_CYCLE_POINT. The value is then stored in suite state dumps and persists across restarts, but it does get wiped out (set to None) after a warm start, because a warm start is really an implicit restart in which all state information is lost (except that the previous cycle is assumed to have completed).
The $CYLC_SUITE_INITIAL_CYCLE_POINT variable allows tasks to determine if they are running in the initial cold-start cycle point, when different behaviour may be required, or in a normal mid-run cycle point. Note however that an initial R1 graph section is now the preferred way to get different behaviour at suite start-up.
Since cylc-4.6.0 any cylc suite can run in live, simulation, or dummy mode. Prior to that release simulation mode was a hybrid mode that replaced real tasks with local dummy tasks. This allowed local simulation testing of any suite, to get the scheduling right without running real tasks, but running dummy tasks locally does not add much value over a pure simulation (in which no tasks are submitted at all) because all job submission configuration has to be ignored and most task job script sections have to be cut out to avoid any code that could potentially be specific to the intended task host. So at 4.6.0 we replaced this with a pure simulation mode (task proxies go through the running state automatically within cylc, and no dummy tasks are submitted to run) and a new dummy mode in which only the real task scripting is dummied out - each dummy task is submitted exactly as the task it represents on the correct host and in the same execution environment. A successful dummy run confirms not only that the scheduling works correctly but also tests real job submission, communication from remote task hosts, and the real task job scripts (in which errors such as use of undefined variables will cause a task to fail).
The run mode, which defaults to live, is set on the command line (for run and restart):
but you can configure the suite to force a particular run mode,
This can be used, for example, for demo suites that necessarily run out of their original context; or to temporarily prevent accidental execution of expensive real tasks during suite development.
Dummy mode task scripting just prints a message and sleeps for ten seconds by default, but you can override this behaviour for particular tasks or task groups if you like. Here’s how to make a task sleep for twenty seconds and then fail in dummy mode:
Finally, in simulation mode each task takes between 1 and 15 seconds to “run” by default, but you can also alter this for particular tasks or groups of tasks:
Note that to get a failed simulation or dummy mode task to succeed on re-triggering, just change the suite.rc file appropriately and reload the suite definition at run time with cylc reload SUITE before re-triggering the task.
Dummy mode is equivalent to removing all user-defined task scripting to expose the default scripting.
The run mode is recorded in the suite state dump file. Cylc will not let you restart a non-live mode suite in live mode, or vice versa - any attempt to do the former would certainly be a mistake (because the simulation mode dummy tasks do not generate any of the real outputs depended on by downstream live tasks), and the latter, while feasible, would corrupt the live state dump by turning it over to simulation mode. The easiest way to test a live suite in simulation mode, if you don’t want to obliterate the current state dump by doing a cold or warm start (as opposed to a restart from the previous state) is to take a quick copy of the suite and run the copy in simulation mode. However, if you really want to run a live suite forward in simulation mode without copying it, do this:
Reference tests are finite-duration suite runs that abort with non-zero exit status if any of the following conditions occur (by default):
The default shutdown event handler for reference tests is cylc hook check-triggering which compares task triggering information (what triggers off what at run time) in the test run suite log to that from an earlier reference run, disregarding the timing and order of events - which can vary according to the external queueing conditions, runahead limit, and so on.
To prepare a reference log for a suite, run it with the --reference-log option, and manually verify the correctness of the reference run.
To reference test a suite, just run it (in dummy mode for the most comprehensive test without running real tasks) with the --reference-test option.
A battery of automated reference tests is used to test cylc before posting a new release version. Reference tests can also be used to check that a cylc upgrade will not break your own complex suites - the triggering check will catch any bug that causes a task to run when it shouldn’t, for instance; even in a dummy mode reference test the full task job script (sans script items) executes on the proper task host by the proper job submission method.
Reference tests can be configured with the following settings:
If the default reference test is not sufficient for your needs, firstly note that you can override the default shutdown event handler, and secondly that the --reference-test option is merely a short cut to the following suite.rc settings which can also be set manually if you wish:
The cylc suite-state command interrogates suite run databases. It has a polling mode that waits for a given task in the target suite to achieve a given state. This can be used to make task scripting wait for a remote task to succeed (for example). The suite graph notation also provides a way to define automatic suite-state polling tasks, which use the same polling command under the hood. Note that cylc suite-state can only trigger off task states in remote suites and does not support triggering off task messages.
Here’s how to trigger a task bar off a task foo in a remote suite called other.suite:
Local task my-foo will poll for the success of foo in suite other.suite, at the same cycle point, succeeding only when or if it succeeds. Other task states can also be polled,
The default polling parameters (e.g. maximum number of polls and the interval between them) are printed by cylc suite-state --help and can be configured if necessary under the local polling task runtime section:
For suites owned by others, or those with run databases in non-standard locations, use the --run-dir option, or in-suite:
If the remote task has a different cycling sequence, just arrange for the local polling task to be on the same sequence as the remote task that it represents. For instance, if local task cat cycles 6-hourly at 0,6,12,18 but needs to trigger off a remote task dog at 3,9,15,21,
For suite-state polling the cycle point of the target task is treated as a literal string so the polling command has to be told if the remote suite has a different cycle point format. Use the --template option for this, or in-suite:
Note that the remote suite does not have to be running when polling commences because the command interrogates the suite run database, not the suite server process.
The following topics have yet to be documented in detail.
Small groups of cylc users can of course share suites by manual copying, and generic revision control tools can be used on cylc suites as for any collection of files. Beyond this cylc does not have a built-in solution for suite storage and discovery, revision control, and deployment, on a network. That is not cylc’s core purpose, and large sites may have preferred revision control systems and suite meta-data requirements that are difficult to anticipate. We can, however, recommend the use of Rose to do all of this very easily and elegantly with cylc suites.
Rose is a framework for managing and running suites of scientific applications, developed at the UK Met Office for use with cylc. It is available under the open source GPL license.
A suite can contain a small number of large, internally complex tasks; a large number of small, simple tasks; or anything in between. Cylc can easily handle a large number of tasks, however, so there are definite advantages to fine-graining:
It should be possible to rerun a task by simply resubmitting it for the same cycle point. In other words, failure at any point during execution of a task should not render a rerun impossible by corrupting the state of some internal-use file, or whatever. It is difficult to overstate the usefulness of being able to rerun the same task multiple times, either outside of the suite with cylc submit, or by re-triggering it within the running suite, when debugging a problem.
If a warm-cycled model uses the exact same file names for its restart files regardless of current cycle point, the only cycle point that can subsequently run successfully is the next one. Instead, restart files should be labelled with current cycle point and maintained in a simple rolling archive. Then you can easily rerun the task for any cycle point still in the archive.
If a task does not depend on files generated by another task then generally speaking it should not trigger off that task in the suite scheduling graph. Unnecessary dependence between tasks restricts functional parallelism at run time, and it makes the suite more difficult to understand. If you need to restrict the number of tasks that are active at once, use runahead limiting (12.12) and internal queues (12.13).
Putting task cycle point in output file or directory names makes archiving and cleanup easier, and it facilitates re-runnability by ensuring that important files do not get overwritten from one cycle to the next.
The cylc cycle-point command computes offsets from a given or current cycle point, and can insert the resulting computed date-time into a filename template string.
Dependence between tasks usually, although not always, take the form of files generated by one task and used by others. It is possible to manage these files across a suite without compromising suite flexibility and portability with hard wired I/O locations.
You may be able to have all tasks, or groups of tasks that need to cooperate, read and write from a common workspace, thereby avoiding the need to explicitly move files around. The suite share directory ($CYLC_SUITE_SHARE_DIR) is provided for this purpose. Similarly, task work directories are private to each task by default but they can be shared to allow multiple tasks to simply read and write from their current working directory. Even if you use other custom I/O directories, define their locations in the suite.rc file rather than hard wiring them into task implementation. Shared workspace locations can be passed to tasks as needed without modifying the task implementation, like this:
Special tasks can be used to move files around, from one task’s output directory to another’s input directory. This should only be necessary across host or filesystem boundaries, however; otherwise simply reference shared locations as shown above.
If your suite contains multiple logically distinct tasks that have similar functionality (e.g. tasks that move files around or generate similar products from different datasets) just have them all call the same underlying command, script, or executable, but provide different input parameters as required.
If all I/O is automatically done in suite-specific locations, such as the under suite share and work directories ($CYLC_SUITE_SHARE_DIRECTORY and $CYLC_TASK_WORK_DIRECTORY), you should be able to run multiple copies of the same suite without interference between them, and other users should be able to copy and run your suites with minimal modification.
Where possible a task should not rely on the action of another task, except for the inputs embodied in the suite dependency graph that it has no choice but to depend on. This makes it as easy as possible to run single tasks alone during suite development and debugging. For example, tasks should create their own output directories if necessary rather than assuming their existence due to the action of another task. Note that if the existing task implementation does not handle output directory creation you can do it in suite pre-script or similar.
Tasks can (of course) run external commands, scripts, and executables; and they can read or otherwise make use of external files. In some cases this may be necessary, but it does leave suites vulnerable to external breakages. Alternatively, suites can be more or less completely self-contained (aside from exposure to network, filesystem, and OS problems) if they have private copies of every file they need at run time. Tasks can access files stored under their suite definition directory via $CYLC_SUITE_DEF_PATH, and the suite bin directory is automatically added to $PATH in the task execution environment. If you have multiple suites there may be a tradeoff between self-containment and duplication of files, but this does not particularly matter if you can automatically extract, build, and install suite files from external repositories prior to, or at the start of, a suite run.
A suite definition and any files stored with it should be version controlled, and a particular revision extracted before a run. The extracted source suite will be a repository clone or working copy, depending on your choice of revision control software, and can be used for further development. The source files should then be installed to another location where the suite will actually be executed (the cylc suite run directory is ideal for this). External files may also be installed into the suite at this time, prior to the run, or by special deployment tasks that run at suite start-up. This makes self-containment easier to achieve, and the clean separation of source and installed suite allows further development without breaking a running suite. Rose (14.1) supports this mode of working with cylc suites.
Correct scheduling is not necessarily equivalent to orderly generation of products in strict date-time order. Under cylc a product generation task will trigger as soon as its private prerequisites are satisfied regardless of whether other tasks at the same cycle point have finished or have yet to run. If your product presentation system demands that all products are uploaded in order, then be aware that this may be quite inefficient if your suite ever has to catch up from a delay or run over historical data, but if necessary you can force tasks to run in the right order even if their true dependencies do not require that. One way to do this is to declare the product upload task to be sequential, which is equivalent to making it depend on its own previous instance (see 9.3.4.13).
Tasks that wait on external real time data should have a clock-trigger to delay submission until roughly the expected time of data availability (see 9.3.4.15), otherwise they could clutter up your batch scheduler queue by submitting hours earlier. Similarly, suite polling tasks (for inter-suite dependence in real time operation) should use a clock-trigger to delay their submission until the expected time of the remote suite event.
Some tasks wait on external events and therefore need to repeatedly check and wait for the event before reporting eventual success (or perhaps failure after a timeout). For example, a task that waits for a file to appear on an ftp server. Typically these should be clock-trigger tasks (see above), but once triggered there are two ways to handle the repeated checking: the task itself could implement a check-and-wait loop; or you could just configure multiple retries for the task (see 12.14).
Cylc suites, without modification, can handle real time and delayed operation equally well. In caught-up real time operation, clock-trigger tasks constrain the behaviour of the whole suite, or at least of any tasks downstream of them in the dependency graph. In delayed or historical operation clock-trigger tasks will not constrain the suite at all, and cylc’s cycle point interleaving abilities come to the fore, because the clock-trigger times have already passed. But if a clock-trigger task catches up to the wall clock, it will automatically wait again. In this way cylc suites naturally transition between delayed and real time operation as required.
To help avoid suite maintenance errors in the future, properties shared by multiple tasks (job submission settings, environment variables, scripting, etc.) should be defined only once, using runtime inheritance (9.4) or Jinja2 variables (9.6).
Multiple inheritance is efficient when tasks share many properties, but Jinja2 variables may be preferred when a small number of properties are shared by tasks that don’t have anything else in common (e.g. a single environment variable for the location of a shared file).
For environment variables in particular it may be tempting to define all variables for all tasks once under [root], but this is analagous to overuse of global variables in programming and it can make it difficult to determine which variables matter to which tasks. Environment filters (A.5.1.23) can be used to make this safer, but generally it is best to provide each task with only the variables that it needs. It is difficult to be sure if a task really needs a variable that is passed to it, but you can be sure that it does not use a variable that is not passed to it.
Finally, Jinja2 can also be used to avoid polluting task environments with variables used for the sole purpose of deriving other variables at task run time. Instead of this:
do this:
If the values of these Jinja2 variables are needed in external scripts, just translate them directly in environment sections:
If you find yourself writing runtime scripting to change a task’s behaviour in some cycle points, consider that the graph is usually the proper place to express this sort of thing. Use different task names, but have them inherit common properties to avoid duplication. Instead of this:
do this:
Similarly, if your task has a different behaviour at the initial or final cycle point, consider using an R1 syntax to separate out the functionality.
Effective visualization can make complex suites easier to understand. Collapsible task families for visualization are defined by the first parents in the runtime namespace hierarchy. Tasks should generally be grouped into visualization families that reflect their purpose within the structure of the suite rather than technical detail such as common job submission method or task host. This often coincides nicely with common configuration inheritance requirements, but if it doesn’t you can use an empty namespace as a first parent for visualization:
and you can demote parents from primary to secondary:
Good style is to some extent a matter of taste. That said, for collaborative development of complex systems it is important to settle on a clear and consistent style, and you may find the following suggestions useful. Note that the boundary between this section (style) and the previous (design) is somewhat arbitrary.
The suite.rc file format consists of item = value pairs under nested section headings. Clear indentation is the best way to show local nesting level inside large blocks.
Don’t align item = value pairs on the = character - this does not show nesting level clearly and it pushes everything off to the right:
The following layout does preserve proper indentation on the left, but the whole block may need reformatting after changing one line, which pollutes your revision history with spurious changes:
is preferred over this (or similar):
The extra whitespace here translates directly to spurious indentation in the task job script. As it happens this is just an aesthetic problem in bash scripts, but for Python job scripts (which cylc may support in the future) it would be a technical error.
Comments should be minimal, but not too minimal. If context and clear item names will do, leave it at that. Extremely verbose comments tend to be neglected and eventually get out of sync with the code, a result that may be worse than having no comments at all.
Keep to the standard maximum line length of 79 characters where possible. Very long lines affect readability, may pose a problem for auto-line-breaking in text editors, and make side-by-side diff display less effective.
Graph lines can also be split up without line breaks, like this:
Use UPPERCASE_NAMES for families and lowercase_names for tasks, so that you can tell which is which at a glance.
Trivial task scripting may be inlined in the suite definition but anything more should be written to a script file. This keeps the suite definition tidy, it allows proper shell-mode text editing, and it allows separate command line testing of the script during development or debugging.
This appendix defines all legal suite definition config items. Embedded Jinja2 code (see 9.6) must process to a valid raw suite.rc file. See also 9.2 for a descriptive overview of suite.rc files, including syntax (9.2.1).
The only top level configuration items at present are the suite title and description.
A single line description of the suite. It is displayed in the db viewer window and can be retrieved at run time with the cylc show command.
A multi-line description of the suite. It can be retrieved by the db viewer right-click menu, or at run time with the cylc show command.
A web URL to suite documentation. If present it can be browsed with the cylc doc command, or from the gcylc Suite menu. The variable ${CYLC_SUITE_NAME} will be replaced with the actual suite name (note this is not a shell environment variable in this context). See also task URLs (A.5.1.4).
]
This section is for configuration that is not specifically task-related.
If this item is set cylc will abort if the suite is not started in the specified mode. This can be used for demo suites that have to be run in simulation mode, for example, because they have been taken out of their normal operational context; or to prevent accidental submission of expensive real tasks during suite development.
Cylc runs off the suite host’s system clock by default. This item allows you to run the suite in UTC even if the system clock is set to local time. Clock-trigger tasks will trigger when the current UTC time is equal to their cycle point date-time plus offset; other time values used, reported, or logged by the suite daemon will usually also be in UTC. The default for this can be set at the site level (see B.14.1).
To just alter the timezone used in the date/time cycle point format, see A.2.5. To just alter the number of expanded year digits (for years below 0 or above 9999), see A.2.4.
Cylc usually uses a CCYYMMDDThhmmZ (Z in the special case of UTC) or CCYYMMDDThhmm+hhmm format (+ standing for + or - here) for writing down date/time cycle points, which follows one of the basic formats outlined in the ISO 8601 standard. For example, a cycle point on the 3rd of February 2001 at 4:50 in the morning, UTC (+0000 timezone), would be written 20010203T0450Z. Similarly, for the the 3rd of February 2001 at 4:50 in the morning, +1300 timezone, cylc would write 20010203T0450+1300.
You may use the isodatetime library’s syntax to write dates and times in ISO 8601 formats - CC for century, YY for decade and decadal year, +X for expanded year digits and their positive or negative sign, thereafter following the ISO 8601 standard example notation except for fractional digits, which are represented as ,ii for hh, ,nn for mm, etc. For example, to write date/times as week dates with fractional hours, set cycle point format to CCYYWwwDThh,iiZ e.g. 1987W041T08,5Z for 08:30 UTC on Monday on the fourth ISO week of 1987.
You can also use a subset of the strptime/strftime POSIX standard - supported tokens are \%F, \%H, \%M, \%S, \%Y, \%d, \%j, \%m, \%s, \%z.
To use the old cylc date-time format (e.g. 2014020106 for 06:00 on the 1st of February 2014), set cycle point format to \%Y\%m\%d\%H.
Please note that using characters like ”/” is not allowed, as it will break task output files (”/” is a reserved character in POSIX file and directory naming). Using ”:” is also not allowed, as it is likely to interfere with usage of commands like ”rsync” when applied to task output files.
For years below 0 or above 9999, the ISO 8601 standard specifies that an extra number of year digits and a sign should be used. This extra number needs to be written down somewhere (here).
For example, if this extra number is set to 2, 00Z on the 1st of January in the year 10040 will be represented as +0100400101T0000Z (2 extra year digits used). With this number set to 3, 06Z on the 4th of May 1985 would be written as +00019850504T0600Z.
This number defaults to 0 (no sign or extra digits used).
If you set UTC mode to True (A.2.2) then this will default to Z. If you use a custom cycle point format (A.2.3), you should specify the timezone choice (or null timezone choice) here as well.
You may set your own time zone choice here, which will be used for all date/time cycle point dumping. Time zones should be expressed as ISO 8601 time zone offsets from UTC, such as +13, +1300, -0500 or +0645, with Z representing the special +0000 case. Cycle points will be converted to the time zone you give and will be represented with this string at the end.
Cycle points that are input without time zones (e.g. as an initial cycle point setting) will use this time zone if set. If this isn’t set (and UTC mode is also not set), then they will default to the current local time zone.
Note that the ISO standard also allows writing the hour and minute separated by a ”:” (e.g. +13:00) - however, this is not recommended, given that the time zone is used as part of task output filenames.
Cylc does not normally abort if tasks fail, but if this item is turned on it will abort with exit status 1 if any task fails.
This has the same effect as the --no-auto-shutdown flag for the suite run commands: it prevents the suite daemon from shutting down normally when all tasks have finished (a suite timeout can still be used to stop the daemon after a period of inactivity, however). This option can make it easier to re-trigger tasks manually near the end of a suite run, during suite development and debugging.
If this is turned on cylc will write the resolved dependencies of each task to the suite log as it becomes ready to run (a list of the IDs of the tasks that actually satisfied its prerequisites at run time). Mainly used for cylc testing and development.
Cylc has internal “hooks” to which you can attach handlers that are called by the suite daemon whenever certain events occur. This section configures suite event hooks; see A.5.1.20 for task event hooks.
Event handler commands can send an email or an SMS, call a pager, intervene in the operation of their own suite, or whatever. They can be held in the suite bin directory, otherwise it is up to you to ensure their location is in $PATH (in the shell in which cylc runs, on the suite host). The commands should require very little resource to run and should return quickly.
Each event handler can be specified as a list of command lines or command line templates.
A command line template may have any or all of these patterns which will be substituted with actual values:
Otherwise, the command line will be called with the following command line arguments:
Additional information can be passed to event handlers via [cylc] →[[environment]].
[cylc] →[[event hooks]] →EVENT handler A comma-separated list of one or more event handlers to call when one of the following EVENTs occurs:
Default values for these can be set at the site level via the siterc file (see B.14.2).
Item details:
[cylc] →[[[event hooks]]] →handlers Specify the general event handlers as a list of command lines or command line templates.
[cylc] →[[[event hooks]]] →handler events Specify the events for which the general event handlers should be invoked.
[cylc] →[[[event hooks]]] →mail events Specify the suite events for which notification emails should be sent.
[cylc] →[[[event hooks]]] →mail from Specify an alternate from: email address for suite event notifications.
[cylc] →[[[events]]] →mail smtp Specify the SMTP server for sending suite event email notifications.
[cylc] →[[[events]]] →mail to A list of email addresses to send suite event notifications. The list can be anything accepted by the mail command.
[cylc] →[[event hooks]] →timeout If a timeout is set and the timeout event is handled, the timeout event handler(s) will be called if the suite times out before it finishes. The timer is set initially at suite start up. It is possible to set a default for this at the site level (see B.14.2).
[cylc] →[[event hooks]] →reset timer If True (the default) the suite timer will continually reset after any task changes state, so you can time out after some interval since the last activity occured rather than on absolute suite execution time.
[cylc] →[[event hooks]] →abort on timeout If a suite timer is set (above) this will cause the suite to abort with error status if the suite times out while still running. It is possible to set a default for this at the site level (see B.14.2).
[cylc] →[[event hooks]] →abort if EVENT handler fails Cylc does not normally care whether an event handler succeeds or fails, but if this is turned on the EVENT handler will be executed in the foreground (which will block the suite while it is running) and the suite will abort if the handler fails.
Environment variables defined in this section are passed to suite and task event handlers.
[cylc] →[[environment]] →__VARIABLE__ Replace __VARIABLE__ with any number of environment variable assignment expressions. Values may refer to other local environment variables (order of definition is preserved) and are not evaluated or manipulated by cylc, so any variable assignment expression that is legal in the shell in which cylc is running can be used (but see the warning above on variable expansions, which will not be evaluated). White space around the ‘=’ is allowed (as far as cylc’s file parser is concerned these are just suite configuration items).
Reference tests are finite-duration suite runs that abort with non-zero exit status if cylc fails, if any task fails, if the suite times out, or if a shutdown event handler that (by default) compares the test run with a reference run reports failure. See 12.22.
[cylc] →[[reference test]] →suite shutdown event handler A shutdown event handler that should compare the test run with the reference run, exiting with zero exit status only if the test run verifies.
As for any event handler, the full path can be ommited if the script is located somewhere in $PATH or in the suite bin directory.
[cylc] →[[reference test]] →required run mode If your reference test is only valid for a particular run mode, this setting will cause cylc to abort if a reference test is attempted in another run mode.
[cylc] →[[reference test]] →allow task failures A reference test run will abort immediately if any task fails, unless this item is set, or a list of expected task failures is provided (below).
[cylc] →[[reference test]] →expected task failures A reference test run will abort immediately if any task fails, unless allow task failures is set (above) or the failed task is found in a list IDs of tasks that are expected to fail.
[cylc] →[[reference test]] →live mode suite timeout The timeout value, expressed as an ISO 8601 duration/interval, after which the test run should be aborted if it has not finished, in live mode. Test runs cannot be done in live mode unless you define a value for this item, because it is not possible to arrive at a sensible default for all suites.
[cylc] →[[reference test]] →simulation mode suite timeout The timeout value in minutes after which the test run should be aborted if it has not finished, in simulation mode. Test runs cannot be done in simulation mode unless you define a value for this item, because it is not possible to arrive at a sensible default for all suites.
[cylc] →[[reference test]] →dummy mode suite timeout The timeout value, expressed as an ISO 8601 duration/interval, after which the test run should be aborted if it has not finished, in dummy mode. Test runs cannot be done in dummy mode unless you define a value for this item, because it is not possible to arrive at a sensible default for all suites.
]
Authentication of client programs with suite daemons can be set in the global site/user config files and overridden here if necessary. See B.15 for more information.
→ public]
The client privilege level granted for public access - i.e. no suite passphrase required. See B.15 for legal values. ]
This section allows cylc to determine when tasks are ready to run.
Cylc runs using the proleptic Gregorian calendar by default. This item allows you to either run the suite using the 360 day calendar (12 months of 30 days in a year) or using integer cycling.
In a cold start each cycling task (unless specifically excluded under [special tasks]) will be loaded into the suite with this cycle point, or with the closest subsequent valid cycle point for the task. Note that special cold-start tasks are not loaded in a warm start. If this item is provided you can override it on the command line or in the gcylc suite start panel.
In date-time cycling, if you do not provide time zone information for this, it will be assumed to be local time, or in UTC if A.2.2 is set, or in the time zone determined by A.2.5 if that is set.
The string “now” converts to the current date-time on the suite host (adjusted to UTC if the suite is in UTC mode but the host is not) to minute resolution. Minutes (or hours, etc.) may be ignored depending on your cycle point format (A.2.3).
Cycling tasks are held once they pass the final cycle point, if one is specified. Once all tasks have achieved this state the suite will shut down. If this item is provided you can override it on the command line or in the gcylc suite start panel.
In date-time cycling, if you do not provide time zone information for this, it will be assumed to be local time, or in UTC if A.2.2 is set, or in the A.2.5 if that is set.
In a cycling suite it is possible to restrict the initial cycle point by defining a list of truncated time points under the initial cycle point constraints.
In a cycling suite it is possible to restrict the final cycle point by defining a list of truncated time points under the final cycle point constraints.
Cycling tasks are held once they pass the hold after cycle point, if one is specified. Unlike the final cycle point suite will not shut down once all tasks have passed this point. If this item is provided you can override it on the command line or in the gcylc suite start panel.
Runahead limiting prevents the fastest tasks in a suite from getting too far ahead of the slowest ones, as documented in 12.12.
This config item specifies a hard limit as a cycle interval between the slowest and fastest tasks. It is deprecated in favour of the newer default limiting by max active cycle points (A.4.8).
Runahead limiting prevents the fastest tasks in a suite from getting too far ahead of the slowest ones, as documented in 12.12.
This config item supersedes the deprecated hard runahead limit (A.4.7). It allows up to N (default 3) consecutive cycle points to be active at any time, adjusted up if necessary for any future triggering.
Configuration of internal queues, by which the number of simultaneously active tasks (submitted or running) can be limited, per queue. By default a single queue called default is defined, with all tasks assigned to it and no limit. To use a single queue for the whole suite just set the limit on the default queue as required. See also 12.13.
[scheduling] →[[queues]] →[[[__QUEUE__]]] Section heading for configuration of a single queue. Replace __QUEUE__ with a queue name, and repeat the section as required.
[scheduling] →[[queues]] →[[[__QUEUE__]]] →limit The maximum number of active tasks allowed at any one time, for this queue.
[scheduling] →[[queues]] →[[[__QUEUE__]]] →members A list of member tasks, or task family names, to assign to this queue (assigned tasks will automatically be removed from the default queue).
This section is used to identify tasks with special behaviour. Family names can be used in special task lists as shorthand for listing all member tasks.
[scheduling] →[[special tasks]] →clock-trigger Clock-trigger tasks (see 9.3.4.15) wait on a wall clock time specified as an offset from their own cycle point.
[scheduling] →[[special tasks]] →clock-expire Clock-expire tasks enter the expired state and skip job submission if too far behind the wall clock when they become ready to run. The expiry time is specified as an offset from wall-clock time; typically it should be negative - see 9.3.4.16.
[scheduling] →[[special tasks]] →external-trigger Externally triggered tasks (see 9.3.4.17) wait on external events reported via the cylc ext-trigger command.
[scheduling] →[[special tasks]] →sequential Sequential tasks are automatically given dependence on their own predecessor. This is equivalent to use of explicit inter-cycle triggers in the graph, except that the automatic version does not show in suite graph visualization. For more on sequential tasks see 9.3.4.13 and 15.4.
[scheduling] →[[special tasks]] →cold-start A cold-start task is one-off task used to satisfy the dependence of an associated task with the same cycle point, on outputs from a previous cycle - when those outputs are not available. The primary use for this is to cold-start a warm-cycled forecast model that normally depends on restart files (e.g. model background fields) generated by its previous forecast, when there is no previous forecast. This is required when cold-starting the suite, but cold-start tasks can also be inserted into a running suite to restart a model that has had to skip some cycles after running into problems. Cold-start tasks can invoke real cold-start processes, or they can just be dummy tasks that represent some external process that has to be completed before the suite is started. Unlike start-up tasks, dependence on cold-start tasks is preseverved in subsequent cycles so they must typically be used in OR’d conditional expressions to avoid holding up the suite.
[scheduling] →[[special tasks]] →one-off Synchronous one-off tasks have an associated cycle point but do not spawn a successor. Synchronous start-up and cold-start tasks are automatically one-off tasks and do not need to be listed here. Dependence on one-off tasks is not restricted to the first cycle.
[scheduling] →[[special tasks]] →exclude at start-up Any task listed here will be excluded from the initial task pool (this goes for suite restarts too). If an inclusion list is also specified, the initial pool will contain only included tasks that have not been excluded. Excluded tasks can still be inserted at run time. Other tasks may still depend on excluded tasks if they have not been removed from the suite dependency graph, in which case some manual triggering, or insertion of excluded tasks, may be required.
[scheduling] →[[special tasks]] →include at start-up If this list is not empty, any task not listed in it will be excluded from the initial task pool (this goes for suite restarts too). If an exclusion list is also specified, the initial pool will contain only included tasks that have not been excluded. Excluded tasks can still be inserted at run time. Other tasks may still depend on excluded tasks if they have not been removed from the suite dependency graph, in which case some manual triggering, or insertion of excluded tasks, may be required.
The suite dependency graph is defined under this section. You can plot the dependency graph as you work on it, with cylc graph or by right clicking on the suite in the db viewer. See also 9.3.
[scheduling] →[[dependencies]] →graph The dependency graph for a completely non-cycling suites can go here. See also A.4.11.2.1 below and 9.3, for graph string syntax.
[scheduling] →[[dependencies]] →[[[__RECURRENCE__]]] __RECURRENCE__ section headings define the sequence of cycle points for which the subsequent graph section is valid. These should be specified in our ISO 8601 derived sequence syntax, or similar for integer cycling:
See 9.3.3 for more on recurrence expressions, and how multiple graph sections combine.
[scheduling] →[[dependencies]] →[[[__RECURRENCE__]]] →graph The dependency graph for a given recurrence section goes here. Syntax examples follow; see also 9.3 and 9.3.4.
]
This section is used to specify how, where, and what to execute when tasks are ready to run. Common configuration can be factored out in a multiple-inheritance hierarchy of runtime namespaces that culminates in the tasks of the suite. Order of precedence is determined by the C3 linearization algorithm as used to find the method resolution order in Python language class hiearchies. For details and examples see 9.4.
Replace __NAME__ with a namespace name, or a comma-separated list of names, and repeat as needed to define all tasks in the suite. Names may contain letters, digits, underscores, and hyphens. A namespace represents a group or family of tasks if other namespaces inherit from it, or a task if no others inherit from it.
If multiple names are listed the subsequent settings apply to each.
All namespaces inherit initially from root, which can be explicitly configured to provide or override default settings for all tasks in the suite.
[runtime] →[[__NAME__]] →inherit A list of the immediate parent(s) this namespace inherits from. If no parents are listed root is assumed.
[runtime] →[[__NAME__]] →title A single line description of this namespace. It is displayed by the cylc list command and can be retrieved from running tasks with the cylc show command.
[runtime] →[[__NAME__]] →description A multi-line description of this namespace, retrievable from running tasks with the cylc show command.
[runtime] →[[__NAME__]] →URL A web URL to task documentation for this suite. If present it can be browsed with the cylc doc command, or by right-clicking on the task in gcylc. The variables ${CYLC_SUITE_NAME} and ${CYLC_TASK_NAME} will be replaced with the actual suite and task names (note that these are not environment variables in this context). See also suite URLs (A.1.3).
(Note that URLs containing the suite comment delimiter # must be protected by quotes).
[runtime] →[[__NAME__]] →init-script This is written to the top of the task job script before the task execution environment is configured, so it does not have access to any suite or task environment variables. It can be a single command or multiple lines of scripting. The original intention was to allow remote tasks to source login scripts to configure their access to cylc, but this should no longer be necessary (see 12.7). See also env-script, pre-script, script, and post-script.
[runtime] →[[__NAME__]] →env-script This is written to the task job script between the cylc-defined environment (suite and task identity, etc.) and the user-defined task runtime environment - i.e. it has access to the cylc environment, and the task environment has access to variables defined by this scripting. It can be a single command or multiple lines of scripting. See also init-script, pre-script, script, and post-script.
[runtime] →[[__NAME__]] →pre-script This is written to the task job script immediately before the script item (just below). It can be a single command or multiple lines of scripting. See also init-script, env-script, script, and post-script.
[runtime] →[[__NAME__]] →script The is the main user-defined scripting to run when the task is ready. It can be a single command or multiple lines of scripting. See also init-script, env-script, pre-script, and post-script.
[runtime] →[[__NAME__]] →post-script This is written to the task job script immediately after the script item (just above). It can be a single command or multiple lines of scripting. See also init-script, env-script, pre-script, and script.
[runtime] →[[__NAME__]] →retry delays A list of ISO 8601 time duration/intervals after which to resubmit the task if it fails. The variable $CYLC_TASK_TRY_NUMBER in the task execution environment is incremented each time, starting from 1 for the first try - this can be used to vary task behavior by try number.
[runtime] →[[__NAME__]] →submission polling intervals A list of intervals, expressed as ISO 8601 duration/intervals, with optional multipliers, after which cylc will poll for status while the task is in the submitted state.
For the polling task communications method this overrides the default submission polling interval in the site/user config files (6). For pyro and ssh task communications polling is not done by default but it can still be configured here as a regular check on the health of submitted tasks.
Each list value is used in turn until the last, which is used repeatedly until finished.
A single interval value is probably appropriate for submission polling.
[runtime] →[[__NAME__]] →execution polling intervals A list of intervals, expressed as ISO 8601 duration/intervals, with optional multipliers, after which cylc will poll for status while the task is in the running state.
For the polling task communications method this overrides the default execution polling interval in the site/user config files (6). For pyro and ssh task communications polling is not done by default but it can still be configured here as a regular check on the health of submitted tasks.
Each list value is used in turn until the last, which is used repeatedly until finished.
[runtime] →[[__NAME__]] →work sub-directory Task job scripts are executed from within work directories created automatically under the suite run directory. A task can get its own work directory from $CYLC_TASK_WORK_DIR (or simply $PWD if it does not cd elsewhere at runtime). The default directory path contains task name and cycle point, to provide a unique workspace for every instance of every task. If several tasks need to exchange files and simply read and write from their from current working directory, this item can be used to override the default to make them all use the same workspace.
The top level share and work directory location can be changed (e.g. to a large data area) by a global config setting (see B.9.1.2).
Note that if you omit cycle point from the work sub-directory path successive instances of the task will share the same workspace. Consider the effect on cycle point offset housekeeping of work directories before doing this.
[runtime] →[[__NAME__]] →enable resurrection If a message is received from a failed task cylc will normally treat this as an error condition, issue a warning, and leave the task in the “failed” state. But if “enable resurrection” is switched on failed tasks can come back from the dead: if the same task job script is executed again cylc will put the task back into the running state and continue as normal when the started message is received. This can be used to handle HPC-style job preemption wherein a resource manager may kill a running task and reschedule it to run again later, to make way for a job with higher immediate priority. See also 12.17
[runtime] →[[__NAME__]] →[[[dummy mode]]] Dummy mode configuration.
[runtime] →[[__NAME__]] →[[[dummy mode]]] →script The main script item for tasks in dummy mode. See A.5.1.8 for documentation.
[runtime] →[[__NAME__]] →[[[dummy mode]]] →disable pre-script This disables the task pre-script in dummy mode.
[runtime] →[[__NAME__]] →[[[dummy mode]]] →disable post-script This disables the task post-script in dummy mode.
[runtime] →[[__NAME__]] →[[[simulation mode]]] Simulation mode configuration.
[runtime] →[[__NAME__]] →[[[simulation mode]]] →run time range This defines a minimum and a maximum duration (expressed as ISO 8601 duration/intervals) which define a range from which the simulation mode task run length will be randomly chosen.
[runtime] →[[__NAME__]] →[[[job submission]]] This section configures the means by which cylc submits task job scripts to run.
[runtime] →[[__NAME__]] →[[[job submission]]] →method See 11 for how job submission works, and how to define new methods. Cylc has a number of built in job submission methods:
[runtime] →[[__NAME__]] →[[[job submission]]] →command template This allows you to override the actual command used by the chosen job submission method. The template’s %(job)s will be substituted by the job file path.
[runtime] →[[__NAME__]] →[[[job submission]]] →shell This is the shell used to interpret the job script submitted by the suite daemon when a task is ready to run. It has no bearing on the shell used in task implementations. Scripting and suite environment variable assignment expressions must be valid for this shell. The latter is currently hardwired into cylc as export item=value - valid for both bash and ksh because value is entirely user-defined - but cylc would have to be modified slightly to allow use of the C shell.
[runtime] →[[__NAME__]] →[[[job submission]]] →retry delays A list of duration (in ISO 8601 syntax), after which to resubmit if job submission fails.
[runtime] →[[__NAME__]] →[[[remote]]] Configure host and username, for tasks that do not run on the suite host account. Passwordless ssh is used to submit the task by the configured job submission method, so you must distribute your ssh key to allow this. Cylc must be installed on remote task hosts, but of the external software dependencies only Pyro is required there (not even that if ssh messaging is used; see below).
[runtime] →[[__NAME__]] →[[[remote]]] →host The remote host for this namespace. This can be a static hostname, an environment variable that holds a hostname, or a command that prints a hostname to stdout. Host selection commands are executed just prior to job submission. The host (static or dynamic) may have an entry in the cylc site or user config file to specify parameters such as the location of cylc on the remote machine; if not, the corresponding local settings (on the suite host) will be assumed to apply on the remote host.
[runtime] →[[__NAME__]] →[[[remote]]] →owner The username of the task host account. This is (only) used in the non-interactive ssh command invoked by the suite daemon to submit the remote task (consequently it may be defined using local environment variables (i.e. the shell in which cylc runs, and [cylc] →[[environment]]).
If you use dynamic host selection and have different usernames on the different selectable hosts, you can configure your $HOME/.ssh/config to handle username translation.
[runtime] →[[__NAME__]] →[[[remote]]] →retrieve job logs Remote task job logs are saved to the suite run directory on the task host, not on the suite host. If you want the job logs pulled back to the suite host automatically, you can set this item to True. The suite will then attempt to rsync the job logs once from the remote host each time a task job completes. E.g. if the job file is ~/cylc-run/tut.oneoff.remote/log/job/1/hello/01/job, anything under ~/cylc-run/tut.oneoff.remote/log/job/1/hello/01/ will be retrieved.
[runtime] →[[__NAME__]] →[[[remote]]] →retrieve job logs max size If the disk space of the suite host is limited, you may want to set the maximum sizes of the job log files to retrieve. The value can be anything that is accepted by the --max-size=SIZE option of the rsync command.
[runtime] →[[__NAME__]] →[[[remote]]] →retrieve job logs retry delays Some batch systems have considerable delays between the time when the job completes and when it writes the job logs in its normal location. If this is the case, you can configure an initial delay and some retry delays between subsequent attempts. The default behaviour is to attempt once without any delay.
[runtime] →[[__NAME__]] →[[[remote]]] →suite definition directory The path to the suite definition directory on the remote host, needed if remote tasks require access to files stored there (via $CYLC_SUITE_DEF_PATH) or in the suite bin directory (via $PATH). If this item is not defined, the local suite definition directory path will be assumed, with the suite owner’s home directory, if present, replaced by '$HOME' for interpretation on the remote host.
[runtime] →[[__NAME__]] →[[[event hooks]]] Cylc can call nominated event handlers when certain task events occur. This section configures specific task event handlers; see A.2.9 for suite event hooks. See A.5.1.21 for general and built-in task event handlers.
Event handler commands can be located in the suite bin/ directory, otherwise it is up to you to ensure their location is in $PATH (in the shell in which cylc runs, on the suite host). The commands should require very little resource to run and should return quickly.
Each task event handler can be specified as a list of command lines or command line templates.
A command line template may have any or all of these patterns which will be substituted with actual values:
Otherwise, the command line will be called with the following arguments:
For an explanation of the substitution syntax, see String Formatting Operations in the Python documentation: https://docs.python.org/2/library/stdtypes.html#string-formatting.
Additional information can be passed to event handlers via the [cylc] →[[environment]] (but not via task runtime environments - event handlers are not called by tasks).
[runtime] →[[__NAME__]] →[[[event hooks]]] →EVENT handler A list of one or more event handlers to call when one of the following EVENTs occurs:
Item details:
[runtime] →[[__NAME__]] →[[[event hooks]]] →submission timeout If a task has not started after the specified ISO 8601 duration/interval, the submission timeout event handler(s) will be called.
[runtime] →[[__NAME__]] →[[[event hooks]]] →execution timeout If a task has not finished after the specified ISO 8601 duration/interval, the execution timeout event handler(s) will be called.
[runtime] →[[__NAME__]] →[[[event hooks]]] →reset timer If you set an execution timeout the timer can be reset to zero every time a message is received from the running task (which indicates the task is still alive). Otherwise, the task will timeout if it does not finish in the alotted time regardless of incoming messages.
[runtime] →[[__NAME__]] →[[[events]]] Cylc can call nominated event handlers when certain task events occur. This section configures general and built-in task event handlers; see A.2.9 for suite event hooks. See A.5.1.20 for specific task event handlers.
Event handler commands can be located in the suite bin/ directory, otherwise it is up to you to ensure their location is in $PATH (in the shell in which cylc runs, on the suite host). The commands should require very little resource to run and should return quickly.
Each task event handler can be specified as a list of command lines or command line templates.
A command line template may have any or all of these patterns which will be substituted with actual values:
Otherwise, the command line will be called with the following command line arguments:
[runtime] →[[__NAME__]] →[[[events]]] →handlers Specify a list of command lines or command line templates as task event handlers.
[runtime] →[[__NAME__]] →[[[events]]] →handler events Specify the events for which the general task event handlers should be invoked.
[runtime] →[[__NAME__]] →[[[events]]] →handler retry delays Specify an initial delay before running an event handler command and any retry delays in case the command returns a non-zero code. The default behaviour is to run an event handler command once without any delay.
[runtime] →[[__NAME__]] →[[[events]]] →mail events Specify the events for which notification emails should be sent.
[runtime] →[[__NAME__]] →[[[events]]] →mail from Specify an alternate from: email address for event notifications.
[runtime] →[[__NAME__]] →[[[events]]] →mail retry delays Specify an initial delay before running the mail notification command and any retry delays in case the command returns a non-zero code. The default behaviour is to run the mail notification command once without any delay.
[runtime] →[[__NAME__]] →[[[events]]] →mail smtp Specify the SMTP server for sending email notifications.
[runtime] →[[__NAME__]] →[[[events]]] →mail to A list of email addresses to send task event notifications. The list can be anything accepted by the mail command.
[runtime] →[[__NAME__]] →[[[events]]] →submission timeout Equivalent to A.5.1.20.2.
[runtime] →[[__NAME__]] →[[[events]]] →execution timeout Equivalent to A.5.1.20.3.
[runtime] →[[__NAME__]] →[[[events]]] →register job logs retry delays Some batch systems have considerable delays between the time when the job completes and when it writes the job logs in its normal location. Consequently, the suite may be unable to register the existence of some job log files in its runtime database until the job log files become available at their expected location. If this is the case, you can configure an initial delay and some retry delays between subsequent attempts. The default behaviour is to attempt once without any delay.
[runtime] →[[__NAME__]] →[[[events]]] →reset timer Equivalent to A.5.1.20.4.
[runtime] →[[__NAME__]] →[[[environment]]] The user defined task execution environment. Variables defined here can refer to cylc suite and task identity variables, which are exported earlier in the task job script, and variable assignment expressions can use cylc utility commands because access to cylc is also configured earlier in the script. See also 9.4.7.
[runtime] →[[__NAME__]] →[[[environment]]] →__VARIABLE__ Replace __VARIABLE__ with any number of environment variable assignment expressions. Order of definition is preserved so values can refer to previously defined variables. Values are passed through to the task job script without evaluation or manipulation by cylc, so any variable assignment expression that is legal in the job submission shell can be used. White space around the ‘=’ is allowed (as far as cylc’s suite.rc parser is concerned these are just normal configuration items).
[runtime] →[[__NAME__]] →[[[environment filter]]] This section contains environment variable inclusion and exclusion lists that can be used to filter the inherited environment. This is not intended as an alternative to a well-designed inheritance hierarchy that provides each task with just the variables it needs. Filters can, however, improve suites with tasks that inherit a lot of environment they don’t need, by making it clear which tasks use which variables. They can optionally be used routinely as explicit “task environment interfaces” too, at some cost to brevity, because they guarantee that variables filtered out of the inherited task environment are not used.
Note that environment filtering is done after inheritance is completely worked out, not at each level on the way, so filter lists in higher-level namespaces only have an effect if they are not overridden by descendants.
[runtime] →[[__NAME__]] →[[[environment filter]]] →include If given, only variables named in this list will be included from the inherited environment, others will be filtered out. Variables may also be explicitly excluded by an exclude list.
[runtime] →[[__NAME__]] →[[[environment filter]]] →exclude Variables named in this list will be filtered out of the inherited environment. Variables may also be implicitly excluded by omission from an include list.
[runtime] →[[__NAME__]] →[[[directives]]] Batch queue scheduler directives. Whether or not these are used depends on the job submission method. For the built-in methods that support directives (loadleveler, lsf, pbs, sge, slurm, moab), directives are written to the top of the task job script in the correct format for the method. Specifying directives individually like this allows use of default directives that can be individually overridden at lower levels of the runtime namespace hierarchy.
[runtime] →[[__NAME__]] →[[[directives]]] →__DIRECTIVE__ Replace __DIRECTIVE__ with each directive assignment, e.g. class = parallel
Example directives for the built-in job submission methods are shown in 11.2.
[runtime] →[[__NAME__]] →[[[outputs]]] Register custom task outputs for use in message triggering in this section (9.3.4.5)
[runtime] →[[__NAME__]] →[[[outputs]]] →__OUTPUT__ Replace __OUTPUT__ with one or more custom task output messages (9.3.4.5). The item name is used to select the custom output message in graph trigger notation.
[runtime] →[[__NAME__]] →[[[suite state polling]]] Configure automatic suite polling tasks as described in 12.23. The items in this section reflect the options and defaults of the cylc suite-state command, except that the target suite name and the --task, --cycle, and --status options are taken from the graph notation.
[runtime] →[[__NAME__]] →[[[suite state polling]]] →run-dir For your own suites the run database location is determined by your site/user config. For other suites, e.g. those owned by others, or mirrored suite databases, use this item to specify the location of the top level cylc run directory (the database should be a suite-name sub-directory of this location).
[runtime] →[[__NAME__]] →[[[suite state polling]]] →template Cycle point template of the target suite, if different from that of the polling suite.
[runtime] →[[__NAME__]] →[[[suite state polling]]] →interval Polling interval expressed as an ISO 8601 duration/interval.
[runtime] →[[__NAME__]] →[[[suite state polling]]] →max-polls The maximum number of polls before timing out and entering the ‘failed’ state.
[runtime] →[[__NAME__]] →[[[suite state polling]]] →user Username of an account on the suite host to which you have access. The polling cylc suite-state command will be invoked on the remote account.
[runtime] →[[__NAME__]] →[[[suite state polling]]] →host The hostname of the target suite. The polling cylc suite-state command will be invoked on the remote account.
[runtime] →[[__NAME__]] →[[[suite state polling]]] →verbose Run the polling cylc suite-state command in verbose output mode.
]
Configuration of suite graphing for the cylc graph command (graph extent, styling, and initial family-collapsed state) and the gcylc graph view (initial family-collapsed state). Graphviz documentation of node shapes and so on can be found at http://www.graphviz.org/Documentation.php.
The initial cycle point for graph plotting.
The visualization initial cycle point gets adjusted up if necessary to the suite initial cycling point.
An explicit final cycle point for graph plotting. If used, this overrides the preferred number of cycle points (below).
The visualization final cycle point gets adjusted down if necessary to the suite final cycle point.
The number of cycle points to graph starting from the visualization initial cycle point. This is the preferred way of defining the graph end point, but it can be overridden by an explicit final cycle point (above).
A list of family (namespace) names to be shown in the collapsed state (i.e. the family members will be replaced by a single family node) when the suite is first plotted in the graph viewer or the gcylc graph view. If this item is not set, the default is to collapse all families at first. Interactive GUI controls can then be used to group and ungroup family nodes at will.
Graph edges (dependency arrows) can be plotted in the same color as the upstream node (task or family) to make paths through a complex graph easier to follow.
Graph node labels can be printed in the same color as the node outline.
Set the default attributes (color and style etc.) of graph nodes (tasks and families). Attribute pairs must be quoted to hide the internal = character.
Set the default attributes (color and style etc.) of graph edges (dependency arrows). Attribute pairs must be quoted to hide the internal = character.
Define named groups of graph nodes (tasks and families) which can styled en masse, by name, in [visualization] →[[node attributes]]. Node groups are automatically defined for all task families, including root, so you can style family and member nodes at once by family name.
[visualization] →[[node groups]] →__GROUP__ Replace __GROUP__ with each named group of tasks or families.
Here you can assign graph node attributes to specific nodes, or to all members of named groups defined in [visualization] →[[node groups]]. Task families are automatically node groups. Styling of a family node applies to all member nodes (tasks and sub-families), but precedence is determined by ordering in the suite definition. For example, if you style a family red and then one of its members green, cylc will plot a red family with one green member; but if you style one member green and then the family red, the red family styling will override the earlier green styling of the member.
[visualization] →[[node attributes]] →__NAME__ Replace __NAME__ with each node or node group for style attribute assignment.
This section defines all legal items and values for cylc site and user config files. See Site And User Config Files (Section 6) for file locations, intended usage, and how to generate the files using the cylc get-site-config command.
As for suite definitions, Jinja2 expressions can be embedded in site and user config files to generate the final result parsed by cylc. Use of Jinja2 in suite definitions is documented in Section 9.6.
A temporary directory is needed by a few cylc commands, and is cleaned automatically on exit. Leave unset for the default (usually $TMPDIR).
Number of process pool worker processes used to execute shell commands (job submission, event handlers, job poll and kill commands).
A rolling archive of suite state dumps is maintained under the suite run directory, and is used for restarts; this item determines the number of previous states retained. The most recent saved state file is called state. Sucessively older files have increasing integer values appended, starting from 1.
Commands that intervene in running suites can be made to ask for confirmation before acting. Some find this annoying and ineffective as a safety measure, however, so command prompts are disabled by default.
The suite run directory tree is created anew with every suite start (not restart) but output from the most recent previous runs can be retained in a rolling archive. Set length to 0 to keep no backups. This is incompatible with current Rose suite housekeeping (see Section 14 for more on Rose) so it is disabled by default, in which case new suite run files will overwrite existing ones in the same run directory tree. Rarely, this can result in incorrect polling results due to the presence of old task status files.
The number of old run directory trees to retain if run directory housekeeping is enabled.
Cylc can poll running jobs to catch problems that prevent task messages from being sent back to the suite, such as hard job kills, network outages, or unplanned task host shutdown. Routine polling is done only for the polling task communication method (below) unless suite-specific polling is configured in the suite definition. A list of interval values can be specified, with the last value used repeatedly until the task is finished - this allows more frequent polling near the beginning and end of the anticipated task run time. Multipliers can be used as shorthand as in the example below.
Cylc can also poll submitted jobs to catch problems that prevent the submitted job from executing at all, such as deletion from an external batch scheduler queue. Routine polling is done only for the polling task communication method (below) unless suite-specific polling is configured in the suite definition. A list of interval values can be specified as for execution polling (above) but a single value is probably sufficient for job submission polling.
When a task host in a suite is a shell command string, cylc calls the shell to determine the task host. This call is invoked by the main process, and may cause the suite to hang while waiting for the command to finish. This setting sets a timeout for such a command to ensure that the suite can continue.
]
This section contains configuration items that affect task-to-suite communications.
If a send fails, the messaging code will retry after a configured delay interval.
If successive sends fail, the messaging code will give up after a configured number of tries.
This is the same as the --pyro-timeout option in cylc commands. Without a timeout Pyro connections to unresponsive suites can hang indefinitely (suites suspended with Ctrl-Z for instance).
]
The suite event log, held under the suite run directory, is maintained as a rolling archive. Logs are rolled over (backed up and started anew) when they reach a configurable limit size.
If true, a new suite log will be started for a new suite run.
How many rolled logs to retain in the archive.
Suite event logs are rolled over when they reach this file size.
]
Documentation locations for the cylc doc command and gcylc Help menus.
File locations of documentation held locally on the cylc host server.
[documentation] →[[files]] →html index File location of the main cylc documentation index.
[documentation] →[[files]] →pdf user guide File location of the cylc User Guide, PDF version.
[documentation] →[[files]] →multi-page html user guide File location of the cylc User Guide, multi-page HTML version.
[documentation] →[[files]] →single-page html user guide File location of the cylc User Guide, single-page HTML version.
Online documentation URLs.
[documentation] →[[urls]] →internet homepage URL of the cylc internet homepage, with links to documentation for the latest official release.
[documentation] →[[urls]] →local index Local intranet URL of the main cylc documentation index.
]
PDF and HTML viewers can be launched by cylc to view the documentation.
Your preferred PDF viewer program.
Your preferred web browser.
]
Choose your favourite text editor for editing suite definitions.
The editor to be invoked by the cylc command line interface.
The editor to be invoked by the cylc GUI.
]
Pyro is the RPC layer used for network communication between cylc clients (suite-connecting commands and guis) servers (running suites). Each suite listens on a dedicated network port, binding on the first available starting at the configured base port.
The first port that cylc is allowed to use.
This determines the maximum number of suites that can run at once on the suite host.
Each suite stores its port number, by suite name, under this directory.
]
Configurable settings for the command line cylc monitor tool.
The sort order for tasks in the monitor view.
]
The [hosts] section configures some important host-specific settings for the suite host (‘localhost’) and remote task hosts. Note that remote task behaviour is determined by the site/user config on the suite host, not on the task host. Suites can specify task hosts that are not listed here, in which case local settings will be assumed, with the local home directory path, if present, replaced by $HOME in items that configure directory locations.
The default task host is the suite host, localhost, with default values as listed below. Use an explicit [hosts][[localhost]] section if you need to override the defaults. Localhost settings are then also used as defaults for other hosts, with the local home directory path replaced as described above. This applies to items omitted from an explicit host section, and to hosts that are not listed at all in the site and user config files. Explicit host sections are only needed if the automatically modified local defaults are not sufficient.
Host section headings can also be regular expressions to match multiple hostnames. Note that the general regular expression wildcard is ‘.⋆’ (zero or more of any character), not ‘⋆’. Hostname matching regular expressions are used as-is in the Python re.match() function. As such they match from the beginning of the hostname string (as specified in the suite definition) and they do not have to match through to the end of the string (use the string-end matching character ‘$’ in the expression to force this).
A hierachy of host match expressions from specific to general can be used because config items are processed in the order specified in the file.
[hosts] →[[HOST]] →run directory The top level of the directory tree that holds suite-specific output logs, state dump files, run database, etc.
[hosts] →[[HOST]] →work directory The top level for suite work and share directories.
[hosts] →[[HOST]] →task communication method The means by which task progress messages are reported back to the running suite. See above for default polling intervals for the poll method.
[hosts] →[[HOST]] →remote copy template A string for the command used to copy files to a remote host. This is not used on the suite host unless you run local tasks under another user account.
[hosts] →[[HOST]] →remote shell template A string for the command used to invoke commands on this host. This is not used on the suite host unless you run local tasks under another user account. (N.B. This setting used to be a string template with a %s being a placeholder for the name of the host. This is no longer the case. Any %s in the string will now be discarded.)
[hosts] →[[HOST]] →use login shell Whether to use a login shell or not for remote command invocation. By default cylc runs remote ssh commands using a login shell,
which will source /etc/profile and ~/.profile to set up the user environment. However, for security reasons some institutions do not allow unattended commands to start login shells, so you can turn off this behaviour to get,
which will use the default shell on the remote machine, sourcing ~/.bashrc (or ~/.cshrc) to set up the environment.
[hosts] →[[HOST]] →cylc executable The cylc executable on a remote host. Note this should point to the cylc multi-version wrapper (see 7.2) on the host, not bin/cylc for a specific installed version. Specify a full path if cylc is not in \$PATH when it is invoked via ssh on this host.
[hosts] →[[HOST]] →global init-script If specified, the value of this setting will be inserted to just before the init-script section of all job scripts that are to be submitted to the specified remote host.
[hosts] →[[HOST]] →copyable environment variables A list containing the names of the environment variables that can and/or need to be copied from the suite daemon to a job.
[hosts] →[[HOST]] →retrieve job logs Global default for the A.5.1.19.3 setting for the specified host.
[hosts] →[[HOST]] →retrieve job logs command If rsync -a is unavailable or insufficient to retrieve job logs from a remote host, you can use this setting to specify a suitable command.
[hosts] →[[HOST]] →retrieve job logs max size Global default for the A.5.1.19.4 setting for the specified host.
[hosts] →[[HOST]] →retrieve job logs retry delays Global default for the A.5.1.19.5 setting for the specified host.
[hosts] →[[HOST]] →task event handler retry delays Host specific default for the A.5.1.21.3 setting.
[hosts] →[[HOST]] →local tail command template A template (with %(filename)s substitution) for the command used to tail-follow local job logs, used by the gcylc log viewer and cylc cat-log --tail. You are unlikely to need to override this.
[hosts] →[[HOST]] →remote tail command template A template (with %(filename)s substitution) for the command used to tail-follow remote job logs, used by the gcylc log viewer and cylc cat-log --tail. The remote tail command needs to be told to die when its parent process exits. You may need to override this command for task hosts where the default tail or ps commands are not equivalent to the Gnu Linux versions.
[hosts] →[[HOST]] →[[[batch systems]]] Settings for particular batch systems on HOST. In the subsections below, SYSTEM should be replaced with the cylc job submission method name that represents the batch system (see A.5.1.18.1).
[hosts] →[[HOST]] →[[[batch systems]]] →[[[[SYSTEM]]]] →err tailer A command template (with %(job_id)s substitution) that can be used to tail-follow the stderr stream of a running job if SYSTEM does not use the normal log file location while the job is running. This setting overrides B.9.1.15 and B.9.1.16 above.
[hosts] →[[HOST]] →[[[batch systems]]] →[[[[SYSTEM]]]] →out tailer A command template (with %(job_id)s substitution) that can be used to tail-follow the stdout stream of a running job if SYSTEM does not use the normal log file location while the job is running. This setting overrides B.9.1.15 and B.9.1.16 above.
[hosts] →[[HOST]] →[[[batch systems]]] →[[[[SYSTEM]]]] →err viewer A command template (with %(job_id)s substitution) that can be used to view the stderr stream of a running job if SYSTEM does not use the normal log file location while the job is running.
[hosts] →[[HOST]] →[[[batch systems]]] →[[[[SYSTEM]]]] →out viewer A command template (with %(job_id)s substitution) that can be used to view the stdout stream of a running job if SYSTEM does not use the normal log file location while the job is running.
]
The suite host’s identity must be determined locally by cylc and passed to running tasks (via $CYLC_SUITE_HOST) so that task messages can target the right suite on the right host.
This item determines how cylc finds the identity of the suite host. For the default name method cylc asks the suite host for its host name. This should resolve on remote task hosts to the IP address of the suite host; if it doesn’t, adjust network settings or use one of the other methods. For the address method, cylc attempts to use a special external “target address” to determine the IP address of the suite host as seen by remote task hosts (in-source documentation in $CYLC_DIR/lib/cylc/suite_host.py explains how this works). And finally, as a last resort, you can choose the hardwired method and manually specify the host name or IP address of the suite host.
This item is required for the address self-identification method. If your suite host sees the internet, a common address such as google.com will do; otherwise choose a host visible on your intranet.
Use this item to explicitly set the name or IP address of the suite host if you have to use the hardwired self-identification method.
]
Utilities such as cylc gsummary need to scan hosts for running suites.
A list of hosts to scan for running suites.
]
Global site/user defaults for A.5.1.21.
]
Settings for the automated development tests.
The name of a remote host with shared HOME file system as the host running the test battery.
The name of a remote host without shared HOME file system as the host running the test battery.
Settings for testing supported batch systems (job submission methods). The tests for a batch system are only performed if the batch system is available on the test host or a remote host accessible via SSH from the test host.
[test battery] →[[batch systems]] →[[[SYSTEM]]] SYSTEM is the name of a supported batch system with automated tests. This can currently be ”loadleveler”, ”lsf”, ”pbs”, ”sge” and/or ”slurm”.
[test battery] →[[batch systems]] →[[[SYSTEM]]] →host The name of a host where commands for this batch system is available. Use ”localhost” if the batch system is available on the host running the test battery. Any specified remote host should be accessible via SSH from the host running the test battery.
[test battery] →[[batch systems]] →[[[SYSTEM]]] →err viewer The command template (with \%(job_id)s substitution) for testing the run time stderr viewer functionality for this batch system.
[test battery] →[[batch systems]] →[[[SYSTEM]]] →out viewer The command template (with \%(job_id)s substitution) for testing the run time stdout viewer functionality for this batch system.
[test battery] →[[batch systems]] →[[[SYSTEM]]] →[[[[directives]]]] The minimum set of directives that must be supplied to the batch system on the site to initiate jobs for the tests.
]
Default values for entries in the suite.rc [cylc] section.
Allows you to set a default value for UTC mode in a suite at the site level. See A.2.2 for details.
You can define site defaults for each of the following options, details of which can be found under A.2.9:
[cylc] →[[event hooks]] →handlers
[cylc] →[[event hooks]] →handler events
[cylc] →[[event hooks]] →startup handler
[cylc] →[[event hooks]] →shutdown handler
[cylc] →[[event hooks]] →mail events
[cylc] →[[event hooks]] →mail from
[cylc] →[[event hooks]] →mail smtp
[cylc] →[[event hooks]] →mail to
[cylc] →[[event hooks]] →timeout handler
[cylc] →[[event hooks]] →timeout
[cylc] →[[event hooks]] →abort on timeout
]
Authentication of client programs with suite daemons can be configured here, and overridden in suites if necessary (see A.3).
The suite-specific passphrase must be installed on a user’s account to authorize full control privileges (see 7.5 and 12.6). In the future we plan to move to a more traditional user account model so that each authorized user can have their own password.
This sets the client privilege level for public access - i.e. no suite passphrase required.
→ hashes]
This sets the hash algorithms used for Pyro HMAC, which are used to authenticate a client (e.g. gcylc) to the suite daemon. The first item is used as the default hash. Subsequent items (if any) are used as fallback hashes (e.g. for backwards compatibility with older suite daemons that are still running with an alternative hash).
The default hashes are SHA-256 followed by MD5. You may want to exclude MD5 for security reasons - e.g. for FIPS (https://en.wikipedia.org/wiki/FIPS_140-2) compliance. Older cylc versions use MD5.
SHA-256 and SHA-512 are the most secure. MD5 is popularly used for HMAC, but is the least secure. SHA-1 is slightly better. Note that HMAC use is not as susceptible to hash collision weakness, so both algorithms are more secure for HMAC use than in the general case.
Clients will attempt to establish a connection using the algorithms from first to last. The exceptions are scanning and pinging, which will use the algorithm configured in the [authentication] →scan hash setting below. If you have lots of suites running with e.g. MD5 hashes, you will want to include ’md5’ in your list of hashes and set it as the default scan hash.
See https://docs.python.org/2/library/hmac.html and https://docs.python.org/2/library/hashlib.html for discussion of hash algorithms.
See [authentication] →hashes.
Configure which of the specified hash algorithms in [authentication] →hashes is used for scanning and pinging (cylc scan, cylc ping).
The default hash algorithm for scans is MD5, for backwards compatibility reasons. You can override it by specifying a different value for this setting (e.g. for FIPS compliance).
The default will become SHA-256 sometime in 2016, when it is unlikely that any old MD5 suites are still around.
This section defines all legal items and values for the gcylc user config file, which should be located in $HOME/.cylc/gcylc.rc. Current settings can be printed with the cylc get-gui-config command.
Set the suite view panel(s) displayed initially, when the GUI starts. This can be changed later using the tool bar.
List suite views, if any, that should be displayed initially in an ungrouped state. Namespace family grouping can be changed later using the tool bar.
Set the task state color theme, common to all views, to use initially. The color theme can be changed later using the tool bar. See gcylc.rc.eg and themes.rc in $CYLC_DIR/conf/gcylcrc/ for how to modify existing color themes or define your own. Use cylc get-gui-config to list your available themes.
Set the suite view panels initial orientation when the GUI starts. This can be changed later using the ”View” menu ”Toggle views side-by-side” option.
Set the size of the task state dot icons displayed in the text and dot views.
If this is not turned off the default sort order for task names and families in the dot and text views will the order they appear in the suite definition. Clicking on the task name column in the treeview will toggle to alphanumeric sort, and a View menu item does the same for the dot view. If turned off, the default sort order is alphanumeric and definition order is not available at all.
The color used to highlight active task filters in gcylc. It must be a name from the X11 rgb.txt file, e.g. SteelBlue; or a quoted hexadecimal color code, e.g. "#ff0000" for red (quotes are required to prevent the hex code being interpreted as a comment).
]
This section may contain task state color theme definitions.
The name of the task state color-theme to be defined in this section.
[themes] →[[THEME]] →inherit You can inherit from another theme in order to avoid defining all states.
[themes] →[[THEME]] →defaults Set default icon attributes for all state icons in this theme.
For the attribute values, COLOR and FONTCOLOR can be color names from the X11 rgb.txt file, e.g. SteelBlue; or hexadecimal color codes, e.g. #ff0000 for red; and STYLE can be “filled” or “unfilled”. See gcylc.rc.eg and themes.rc in $CYLC_DIR/conf/gcylcrc/ for examples.
[themes] →[[THEME]] →STATE Set icon attributes for all task states in THEME, or for a subset of them if you have used theme inheritance and/or defaults. Legal values of STATE are any of the cylc task proxy states: waiting, runahead, held, queued, ready, submitted, submit-failed, running, succeeded, failed, retrying, submit-retrying.
For the attribute values, COLOR and FONTCOLOR can be color names from the X11 rgb.txt file, e.g. SteelBlue; or hexadecimal color codes, e.g. #ff0000 for red; and STYLE can be “filled” or “unfilled”. See gcylc.rc.eg and themes.rc in $CYLC_DIR/conf/gcylcrc/ for examples.
The graph view in the gcylc GUI shows the structure of the suite as it evolves. It can work well even for large suites, but be aware that the graphviz layout engine has to do a new global layout every time a task proxy appears in or disappears from the task pool. The following may help mitigate any jumping layout problems:
Pyro (Python Remote Objects) is a widely used open source objected oriented Remote Procedure Call technology developed by Irmen de Jong.
Earlier versions of cylc used the Pyro Nameserver to marshal communication between client programs (tasks, commands, viewers, etc.) and their target suites. This worked well, but in principle it provided a route for one suite or user on the network to bring down all running suites by killing the nameserver. Consequently cylc now uses Pyro simply as a lightweight object oriented wrapper for direct network socket communication between client programs and their target suites - all suites are thus entirely isolated from one another.
Cylc 6 introduced new date-time-related syntax for the suite.rc file. In some places, this is quite radically different from the earlier syntax.
Timeouts and delays such as [cylc][[event hooks]]timeout or [runtime][[my_task]]retry delays were written in a purely numeric form before cylc 6, in seconds, minutes (most common), or hours, depending on the setting.
They are now written in an ISO 8601 duration form, which has the benefit that the units are user-selectable (use 1 day instead of 1440 minutes) and explicit.
Nearly all timeouts and delays in cylc were in minutes, except for:
[runtime][[my_task]][[[suite state polling]]]interval
[runtime][[my_task]][[[simulation mode]]]run time range
which were in seconds, and
[scheduling]runahead limit
which was in hours (this is a special case discussed below in J.2).
See Table 1.
| Setting | Pre-Cylc-6 | Cylc-6+ |
[cylc][[event hooks]]timeout | 180 | PT3H |
[runtime][[my_task]]retry delays | 2*30, 360, | 2*PT30M, PT6H, |
| 1440 | P1D | |
[runtime][[my_task]][[[suite state polling]]]interval | 2 | PT2S |
See A.4.7.
The [scheduling]runahead limit setting was written as a number of hours in pre-cylc-6 suites. This is now in ISO 8601 format for date-time cycling suites, so [scheduling]runahead limit=36 would be written [scheduling]runahead limit=PT36H.
There is a new preferred alternative to runahead limit, [scheduling]max active cycle points. This allows the user to configure how many cycle points can run at once (default 3). See A.4.8.
See A.4.2.
The following suite.rc settings have changed name (Table 2):
| Pre-Cylc-6 | Cylc-6+ |
[scheduling]initial cycle time | [scheduling]initial cycle point |
[scheduling]final cycle time | [scheduling]final cycle point |
[visualization]initial cycle time | [visualization]initial cycle point |
[visualization]final cycle time | [visualization]final cycle point |
This change is to reflect the fact that cycling in cylc 6+ can now be over e.g. integers instead of being purely based on date-time.
Date-times written in initial cycle time and final cycle time were in a cylc-specific 10-digit (or less) CCYYMMDDhh format, such as 2014021400 for 00:00 on the 14th of February 2014.
Date-times are now required to be ISO 8601 compatible. This can be achieved easily enough by inserting a T between the day and the hour digits.
| Setting | Pre-Cylc-6 | Cylc-6+ |
[scheduling]initial cycle time | 2014021400 | 20140214T00 |
Start-up tasks have been removed from cylc 6, and use of cold-start tasks is no longer recommended. Instead, you should use the initial/repeat-once notation as detailed in 7.23.2 and 9.3.3.2.3.
Repeating asynchronous tasks have also been removed because non date-time workflows can now be handled more easily with integer cycling. See for instance the satellite data processing example documented in 9.3.3.2.4.
For repeating tasks with hour-based cycling the syntax has only minor changes:
Pre-cylc-6:
Hour-based cycling section names are easy enough to convert, as seen in Table 4.
| Pre-Cylc-6 | Cylc-6+ |
[scheduling][[dependencies]][[[0]]] | [scheduling][[dependencies]][[[T00]]] |
[scheduling][[dependencies]][[[6]]] | [scheduling][[dependencies]][[[T06]]] |
[scheduling][[dependencies]][[[12]]] | [scheduling][[dependencies]][[[T12]]] |
[scheduling][[dependencies]][[[18]]] | [scheduling][[dependencies]][[[T18]]] |
The graph text in hour-based cycling is also easy to convert, as seen in Table 5.
| Pre-Cylc-6 | Cylc-6+ |
my_task[T-6] | my_task[-PT6H] |
my_task[T-12] | my_task[-PT12H] |
my_task[T-24] | my_task[-PT24H] or evenmy_task[-P1D] |
Prior to cylc-6 inter-cycle triggers implicitly created task instances at the offset cycle points. For example, this pre cylc-6 suite automatically creates instances of task foo at the offset hours 3,9,15,21 each day, for task bar to trigger off at 0,6,12,18:
You can run this suite to see how it works (cylc-6+ is backward-compatible: it generates the old-style behaviour when it detects an old-style suite definition). Here’s the direct translation of this suite to cylc-6+ format:
If you run this suite, bar.20140808T00 will execute at start-up because dependence on tasks prior to the initial cycle point is ignored, but the suite will then stall with bar.20140808T06 waiting on foo.20140808T03 (right-click on the task in the GUI to see its prerequisites), which does not exist because the offset foo instances have not been created. Note that if you graph the suite the offset foo instances do appear. That’s because the graph expresses dependence, and bar really does depend on these instances of foo. The problem is just that the tasks don’t get created at run time, which is why the suite stalls. Here’s the correct way to get the desired behaviour in cylc-6+:
Implicit creation of task instances by offset triggers has been disabled because it is error prone: if a trigger offset is wrong it should cause a triggering failure rather than create task instances at incorrect cycle points.
Copyright Ⓒ 2007 Free Software Foundation, Inc. http://fsf.org/
Everyone is permitted to copy and distribute verbatim copies of this
license document, but changing it is not allowed.
Preamble
The GNU General Public License is a free, copyleft license for software and other kinds of works.
The licenses for most software and other practical works are designed to take away your freedom to share and change the works. By contrast, the GNU General Public License is intended to guarantee your freedom to share and change all versions of a program–to make sure it remains free software for all its users. We, the Free Software Foundation, use the GNU General Public License for most of our software; it applies also to any other work released this way by its authors. You can apply it to your programs, too.
When we speak of free software, we are referring to freedom, not price. Our General Public Licenses are designed to make sure that you have the freedom to distribute copies of free software (and charge for them if you wish), that you receive source code or can get it if you want it, that you can change the software or use pieces of it in new free programs, and that you know you can do these things.
To protect your rights, we need to prevent others from denying you these rights or asking you to surrender the rights. Therefore, you have certain responsibilities if you distribute copies of the software, or if you modify it: responsibilities to respect the freedom of others.
For example, if you distribute copies of such a program, whether gratis or for a fee, you must pass on to the recipients the same freedoms that you received. You must make sure that they, too, receive or can get the source code. And you must show them these terms so they know their rights.
Developers that use the GNU GPL protect your rights with two steps: (1) assert copyright on the software, and (2) offer you this License giving you legal permission to copy, distribute and/or modify it.
For the developers’ and authors’ protection, the GPL clearly explains that there is no warranty for this free software. For both users’ and authors’ sake, the GPL requires that modified versions be marked as changed, so that their problems will not be attributed erroneously to authors of previous versions.
Some devices are designed to deny users access to install or run modified versions of the software inside them, although the manufacturer can do so. This is fundamentally incompatible with the aim of protecting users’ freedom to change the software. The systematic pattern of such abuse occurs in the area of products for individuals to use, which is precisely where it is most unacceptable. Therefore, we have designed this version of the GPL to prohibit the practice for those products. If such problems arise substantially in other domains, we stand ready to extend this provision to those domains in future versions of the GPL, as needed to protect the freedom of users.
Finally, every program is threatened constantly by software patents. States should not allow patents to restrict development and use of software on general-purpose computers, but in those that do, we wish to avoid the special danger that patents applied to a free program could make it effectively proprietary. To prevent this, the GPL assures that patents cannot be used to render the program non-free.
The precise terms and conditions for copying, distribution and modification follow.
Terms and Conditions
“This License” refers to version 3 of the GNU General Public License.
“Copyright” also means copyright-like laws that apply to other kinds of works, such as semiconductor masks.
“The Program” refers to any copyrightable work licensed under this License. Each licensee is addressed as “you”. “Licensees” and “recipients” may be individuals or organizations.
To “modify” a work means to copy from or adapt all or part of the work in a fashion requiring copyright permission, other than the making of an exact copy. The resulting work is called a “modified version” of the earlier work or a work “based on” the earlier work.
A “covered work” means either the unmodified Program or a work based on the Program.
To “propagate” a work means to do anything with it that, without permission, would make you directly or secondarily liable for infringement under applicable copyright law, except executing it on a computer or modifying a private copy. Propagation includes copying, distribution (with or without modification), making available to the public, and in some countries other activities as well.
To “convey” a work means any kind of propagation that enables other parties to make or receive copies. Mere interaction with a user through a computer network, with no transfer of a copy, is not conveying.
An interactive user interface displays “Appropriate Legal Notices” to the extent that it includes a convenient and prominently visible feature that (1) displays an appropriate copyright notice, and (2) tells the user that there is no warranty for the work (except to the extent that warranties are provided), that licensees may convey the work under this License, and how to view a copy of this License. If the interface presents a list of user commands or options, such as a menu, a prominent item in the list meets this criterion.
The “source code” for a work means the preferred form of the work for making modifications to it. “Object code” means any non-source form of a work.
A “Standard Interface” means an interface that either is an official standard defined by a recognized standards body, or, in the case of interfaces specified for a particular programming language, one that is widely used among developers working in that language.
The “System Libraries” of an executable work include anything, other than the work as a whole, that (a) is included in the normal form of packaging a Major Component, but which is not part of that Major Component, and (b) serves only to enable use of the work with that Major Component, or to implement a Standard Interface for which an implementation is available to the public in source code form. A “Major Component”, in this context, means a major essential component (kernel, window system, and so on) of the specific operating system (if any) on which the executable work runs, or a compiler used to produce the work, or an object code interpreter used to run it.
The “Corresponding Source” for a work in object code form means all the source code needed to generate, install, and (for an executable work) run the object code and to modify the work, including scripts to control those activities. However, it does not include the work’s System Libraries, or general-purpose tools or generally available free programs which are used unmodified in performing those activities but which are not part of the work. For example, Corresponding Source includes interface definition files associated with source files for the work, and the source code for shared libraries and dynamically linked subprograms that the work is specifically designed to require, such as by intimate data communication or control flow between those subprograms and other parts of the work.
The Corresponding Source need not include anything that users can regenerate automatically from other parts of the Corresponding Source.
The Corresponding Source for a work in source code form is that same work.
All rights granted under this License are granted for the term of copyright on the Program, and are irrevocable provided the stated conditions are met. This License explicitly affirms your unlimited permission to run the unmodified Program. The output from running a covered work is covered by this License only if the output, given its content, constitutes a covered work. This License acknowledges your rights of fair use or other equivalent, as provided by copyright law.
You may make, run and propagate covered works that you do not convey, without conditions so long as your license otherwise remains in force. You may convey covered works to others for the sole purpose of having them make modifications exclusively for you, or provide you with facilities for running those works, provided that you comply with the terms of this License in conveying all material for which you do not control copyright. Those thus making or running the covered works for you must do so exclusively on your behalf, under your direction and control, on terms that prohibit them from making any copies of your copyrighted material outside their relationship with you.
Conveying under any other circumstances is permitted solely under the conditions stated below. Sublicensing is not allowed; section 10 makes it unnecessary.
No covered work shall be deemed part of an effective technological measure under any applicable law fulfilling obligations under article 11 of the WIPO copyright treaty adopted on 20 December 1996, or similar laws prohibiting or restricting circumvention of such measures.
When you convey a covered work, you waive any legal power to forbid circumvention of technological measures to the extent such circumvention is effected by exercising rights under this License with respect to the covered work, and you disclaim any intention to limit operation or modification of the work as a means of enforcing, against the work’s users, your or third parties’ legal rights to forbid circumvention of technological measures.
You may convey verbatim copies of the Program’s source code as you receive it, in any medium, provided that you conspicuously and appropriately publish on each copy an appropriate copyright notice; keep intact all notices stating that this License and any non-permissive terms added in accord with section 7 apply to the code; keep intact all notices of the absence of any warranty; and give all recipients a copy of this License along with the Program.
You may charge any price or no price for each copy that you convey, and you may offer support or warranty protection for a fee.
You may convey a work based on the Program, or the modifications to produce it from the Program, in the form of source code under the terms of section 4, provided that you also meet all of these conditions:
A compilation of a covered work with other separate and independent works, which are not by their nature extensions of the covered work, and which are not combined with it such as to form a larger program, in or on a volume of a storage or distribution medium, is called an “aggregate” if the compilation and its resulting copyright are not used to limit the access or legal rights of the compilation’s users beyond what the individual works permit. Inclusion of a covered work in an aggregate does not cause this License to apply to the other parts of the aggregate.
You may convey a covered work in object code form under the terms of sections 4 and 5, provided that you also convey the machine-readable Corresponding Source under the terms of this License, in one of these ways:
A separable portion of the object code, whose source code is excluded from the Corresponding Source as a System Library, need not be included in conveying the object code work.
A “User Product” is either (1) a “consumer product”, which means any tangible personal property which is normally used for personal, family, or household purposes, or (2) anything designed or sold for incorporation into a dwelling. In determining whether a product is a consumer product, doubtful cases shall be resolved in favor of coverage. For a particular product received by a particular user, “normally used” refers to a typical or common use of that class of product, regardless of the status of the particular user or of the way in which the particular user actually uses, or expects or is expected to use, the product. A product is a consumer product regardless of whether the product has substantial commercial, industrial or non-consumer uses, unless such uses represent the only significant mode of use of the product.
“Installation Information” for a User Product means any methods, procedures, authorization keys, or other information required to install and execute modified versions of a covered work in that User Product from a modified version of its Corresponding Source. The information must suffice to ensure that the continued functioning of the modified object code is in no case prevented or interfered with solely because modification has been made.
If you convey an object code work under this section in, or with, or specifically for use in, a User Product, and the conveying occurs as part of a transaction in which the right of possession and use of the User Product is transferred to the recipient in perpetuity or for a fixed term (regardless of how the transaction is characterized), the Corresponding Source conveyed under this section must be accompanied by the Installation Information. But this requirement does not apply if neither you nor any third party retains the ability to install modified object code on the User Product (for example, the work has been installed in ROM).
The requirement to provide Installation Information does not include a requirement to continue to provide support service, warranty, or updates for a work that has been modified or installed by the recipient, or for the User Product in which it has been modified or installed. Access to a network may be denied when the modification itself materially and adversely affects the operation of the network or violates the rules and protocols for communication across the network.
Corresponding Source conveyed, and Installation Information provided, in accord with this section must be in a format that is publicly documented (and with an implementation available to the public in source code form), and must require no special password or key for unpacking, reading or copying.
“Additional permissions” are terms that supplement the terms of this License by making exceptions from one or more of its conditions. Additional permissions that are applicable to the entire Program shall be treated as though they were included in this License, to the extent that they are valid under applicable law. If additional permissions apply only to part of the Program, that part may be used separately under those permissions, but the entire Program remains governed by this License without regard to the additional permissions.
When you convey a copy of a covered work, you may at your option remove any additional permissions from that copy, or from any part of it. (Additional permissions may be written to require their own removal in certain cases when you modify the work.) You may place additional permissions on material, added by you to a covered work, for which you have or can give appropriate copyright permission.
Notwithstanding any other provision of this License, for material you add to a covered work, you may (if authorized by the copyright holders of that material) supplement the terms of this License with terms:
All other non-permissive additional terms are considered “further restrictions” within the meaning of section 10. If the Program as you received it, or any part of it, contains a notice stating that it is governed by this License along with a term that is a further restriction, you may remove that term. If a license document contains a further restriction but permits relicensing or conveying under this License, you may add to a covered work material governed by the terms of that license document, provided that the further restriction does not survive such relicensing or conveying.
If you add terms to a covered work in accord with this section, you must place, in the relevant source files, a statement of the additional terms that apply to those files, or a notice indicating where to find the applicable terms.
Additional terms, permissive or non-permissive, may be stated in the form of a separately written license, or stated as exceptions; the above requirements apply either way.
You may not propagate or modify a covered work except as expressly provided under this License. Any attempt otherwise to propagate or modify it is void, and will automatically terminate your rights under this License (including any patent licenses granted under the third paragraph of section 11).
However, if you cease all violation of this License, then your license from a particular copyright holder is reinstated (a) provisionally, unless and until the copyright holder explicitly and finally terminates your license, and (b) permanently, if the copyright holder fails to notify you of the violation by some reasonable means prior to 60 days after the cessation.
Moreover, your license from a particular copyright holder is reinstated permanently if the copyright holder notifies you of the violation by some reasonable means, this is the first time you have received notice of violation of this License (for any work) from that copyright holder, and you cure the violation prior to 30 days after your receipt of the notice.
Termination of your rights under this section does not terminate the licenses of parties who have received copies or rights from you under this License. If your rights have been terminated and not permanently reinstated, you do not qualify to receive new licenses for the same material under section 10.
You are not required to accept this License in order to receive or run a copy of the Program. Ancillary propagation of a covered work occurring solely as a consequence of using peer-to-peer transmission to receive a copy likewise does not require acceptance. However, nothing other than this License grants you permission to propagate or modify any covered work. These actions infringe copyright if you do not accept this License. Therefore, by modifying or propagating a covered work, you indicate your acceptance of this License to do so.
Each time you convey a covered work, the recipient automatically receives a license from the original licensors, to run, modify and propagate that work, subject to this License. You are not responsible for enforcing compliance by third parties with this License.
An “entity transaction” is a transaction transferring control of an organization, or substantially all assets of one, or subdividing an organization, or merging organizations. If propagation of a covered work results from an entity transaction, each party to that transaction who receives a copy of the work also receives whatever licenses to the work the party’s predecessor in interest had or could give under the previous paragraph, plus a right to possession of the Corresponding Source of the work from the predecessor in interest, if the predecessor has it or can get it with reasonable efforts.
You may not impose any further restrictions on the exercise of the rights granted or affirmed under this License. For example, you may not impose a license fee, royalty, or other charge for exercise of rights granted under this License, and you may not initiate litigation (including a cross-claim or counterclaim in a lawsuit) alleging that any patent claim is infringed by making, using, selling, offering for sale, or importing the Program or any portion of it.
A “contributor” is a copyright holder who authorizes use under this License of the Program or a work on which the Program is based. The work thus licensed is called the contributor’s “contributor version”.
A contributor’s “essential patent claims” are all patent claims owned or controlled by the contributor, whether already acquired or hereafter acquired, that would be infringed by some manner, permitted by this License, of making, using, or selling its contributor version, but do not include claims that would be infringed only as a consequence of further modification of the contributor version. For purposes of this definition, “control” includes the right to grant patent sublicenses in a manner consistent with the requirements of this License.
Each contributor grants you a non-exclusive, worldwide, royalty-free patent license under the contributor’s essential patent claims, to make, use, sell, offer for sale, import and otherwise run, modify and propagate the contents of its contributor version.
In the following three paragraphs, a “patent license” is any express agreement or commitment, however denominated, not to enforce a patent (such as an express permission to practice a patent or covenant not to sue for patent infringement). To “grant” such a patent license to a party means to make such an agreement or commitment not to enforce a patent against the party.
If you convey a covered work, knowingly relying on a patent license, and the Corresponding Source of the work is not available for anyone to copy, free of charge and under the terms of this License, through a publicly available network server or other readily accessible means, then you must either (1) cause the Corresponding Source to be so available, or (2) arrange to deprive yourself of the benefit of the patent license for this particular work, or (3) arrange, in a manner consistent with the requirements of this License, to extend the patent license to downstream recipients. “Knowingly relying” means you have actual knowledge that, but for the patent license, your conveying the covered work in a country, or your recipient’s use of the covered work in a country, would infringe one or more identifiable patents in that country that you have reason to believe are valid.
If, pursuant to or in connection with a single transaction or arrangement, you convey, or propagate by procuring conveyance of, a covered work, and grant a patent license to some of the parties receiving the covered work authorizing them to use, propagate, modify or convey a specific copy of the covered work, then the patent license you grant is automatically extended to all recipients of the covered work and works based on it.
A patent license is “discriminatory” if it does not include within the scope of its coverage, prohibits the exercise of, or is conditioned on the non-exercise of one or more of the rights that are specifically granted under this License. You may not convey a covered work if you are a party to an arrangement with a third party that is in the business of distributing software, under which you make payment to the third party based on the extent of your activity of conveying the work, and under which the third party grants, to any of the parties who would receive the covered work from you, a discriminatory patent license (a) in connection with copies of the covered work conveyed by you (or copies made from those copies), or (b) primarily for and in connection with specific products or compilations that contain the covered work, unless you entered into that arrangement, or that patent license was granted, prior to 28 March 2007.
Nothing in this License shall be construed as excluding or limiting any implied license or other defenses to infringement that may otherwise be available to you under applicable patent law.
If conditions are imposed on you (whether by court order, agreement or otherwise) that contradict the conditions of this License, they do not excuse you from the conditions of this License. If you cannot convey a covered work so as to satisfy simultaneously your obligations under this License and any other pertinent obligations, then as a consequence you may not convey it at all. For example, if you agree to terms that obligate you to collect a royalty for further conveying from those to whom you convey the Program, the only way you could satisfy both those terms and this License would be to refrain entirely from conveying the Program.
Notwithstanding any other provision of this License, you have permission to link or combine any covered work with a work licensed under version 3 of the GNU Affero General Public License into a single combined work, and to convey the resulting work. The terms of this License will continue to apply to the part which is the covered work, but the special requirements of the GNU Affero General Public License, section 13, concerning interaction through a network will apply to the combination as such.
The Free Software Foundation may publish revised and/or new versions of the GNU General Public License from time to time. Such new versions will be similar in spirit to the present version, but may differ in detail to address new problems or concerns.
Each version is given a distinguishing version number. If the Program specifies that a certain numbered version of the GNU General Public License “or any later version” applies to it, you have the option of following the terms and conditions either of that numbered version or of any later version published by the Free Software Foundation. If the Program does not specify a version number of the GNU General Public License, you may choose any version ever published by the Free Software Foundation.
If the Program specifies that a proxy can decide which future versions of the GNU General Public License can be used, that proxy’s public statement of acceptance of a version permanently authorizes you to choose that version for the Program.
Later license versions may give you additional or different permissions. However, no additional obligations are imposed on any author or copyright holder as a result of your choosing to follow a later version.
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM “AS IS” WITHOUT WARRANTY OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.
If the disclaimer of warranty and limitation of liability provided above cannot be given local legal effect according to their terms, reviewing courts shall apply local law that most closely approximates an absolute waiver of all civil liability in connection with the Program, unless a warranty or assumption of liability accompanies a copy of the Program in return for a fee.
End of Terms and Conditions
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest possible use to the public, the best way to achieve this is to make it free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest to attach them to the start of each source file to most effectively state the exclusion of warranty; and each file should have at least the “copyright” line and a pointer to where the full notice is found.
Also add information on how to contact you by electronic and paper mail.
If the program does terminal interaction, make it output a short notice like this when it starts in an interactive mode:
The hypothetical commands show w and show c should show the appropriate parts of the General Public License. Of course, your program’s commands might be different; for a GUI interface, you would use an “about box”.
You should also get your employer (if you work as a programmer) or school, if any, to sign a “copyright disclaimer” for the program, if necessary. For more information on this, and how to apply and follow the GNU GPL, see http://www.gnu.org/licenses/.
The GNU General Public License does not permit incorporating your program into proprietary programs. If your program is a subroutine library, you may consider it more useful to permit linking proprietary applications with the library. If this is what you want to do, use the GNU Lesser General Public License instead of this License. But first, please read http://www.gnu.org/philosophy/why-not-lgpl.html.
1Future plans for EcoConnect include additional deterministic regional weather forecasts and a statistical ensemble.
2An OR operator on the right doesn’t make much sense: if “B or C” triggers off A, what exactly should cylc do when A finishes?
3In NWP forecast analysis suites parts of the observation processing and data assimilation subsystem will typically also depend on model background fields generated by the previous forecast.
4If you accidentally delete a port file while a suite is running, use cylc scan to determine the port number then use it on the command line (--port) or rewrite the port file manually.
